Peregrine Connects Existing City Data to Power 400 Public Safety Agencies

Peregrine连接现有城市数据，赋能400家公共安全机构 Peregrine通过整合碎片化的市政数据而非部署新的监控硬件，在保障数据主权的同时，使长时程AI智能体能够侦破悬案并追踪本地威胁。 Peregrine联合创始人Nick Noone和Ben Rudolph阐述了其平台如何在维护本地数据主权的前提下，连接并治理涵盖警察、消防、EMS及卫生部门的现有市政数据。他们分享了公司的创立历程——从经历二十多次被拒，到与San Pablo Police Department实现早期突破——并展示了前置部署工程与长时程AI智能体如何解决复杂的现实世界问题。典型案例包括：悬案智能体从300GB证据中提取关键蜂窝网络定位记录、查明水上救援频发的根本原因，以及识别针对当地犹太会堂的协同威胁。最后，他们强调秉持非政治化的“沉稳专业”方式，为全球超过10,000个社区推进技术基础设施现代化。 与依靠摄像头和车牌识别（ALPR）等硬件扩张的传统厂商不同，Peregrine 颠覆了公共安全数据模式，坚持零硬件、不收集新数据，专注于打通并治理城市已有的分散数据资产。 (18:16) 在遭遇二十多次拒绝后，Nick Noone 与 Ben Rudolph 于 2018 年 2 月获得加州圣巴勃罗警察局（San Pablo PD）准入，确立了以深度同理心与一线共担问题的驻场工程（FDE）模式。 (8:25) Peregrine 内部约 90% 的数据集成代码由 AI Agent 自动生成，其研发的悬案 Agent 能够自主深度推理 200 至 300 GB 的多模态案件卷宗，不仅复现了人工平反冤案的证据链，还协助威斯康星州警方从海量通话记录中锁定关键定位证据。 (31:45) 公司推行“反网络效应”与本地数据主权架构，避免构建中心化监控全景系统，并通过垂直整合底层技术栈将部署年费降至 100 万美元以下，服务于更广泛的中小型市政机构。 (48:50)

Peregrine co-founders Nick Noone and Ben Rudolph argue that municipal safety and civic flourishing depend on integrating existing public data rather than constructing centralized surveillance networks. By pairing high-empathy forward deployed engineering with internal AI agents, the company aims to provide local governments with advanced analytical capabilities while strictly preserving municipal data sovereignty. Inverting Surveillance Through Municipal Data Sovereignty (18:16) Contrasting with legacy public safety vendors that rely on proprietary hardware like automated license plate readers and cameras to capture new data into proprietary silos, Peregrine operates on a zero-sensor, zero-new-collection model. The platform acts solely as an integration, structuring, and governance layer over fragmented data that municipal agencies already own. Noone and Rudolph explicitly reject traditional Silicon Valley network effects, arguing that centralizing public safety data into a single repository creates an authoritarian surveillance risk that rightfully alarms the public. To address civil liberties concerns and public skepticism of AI, Peregrine's architecture enforces strict data sovereignty and fine-grained permissioning. Local agencies retain absolute ownership of their data rather than pooling it into an external hub. According to Rudolph, providing secure, auditable digital sharing mechanisms eliminates unsafe legacy practices, such as physically transporting hard drives across jurisdictions, while ensuring municipal operations comply with local privacy ordinances. Forward Deployment as Research and Product Development (1:28) Drawing from Noone's background leading Palantir's SOCOM unit and Rudolph's humanitarian logistics work with the UN Refugee Agency and Dimagi, Peregrine established a deployment model centered on institutional humility. The founders experienced over two dozen agency rejections before Commander Brian Bubar granted them operational access at the San Pablo Police Department in February 2018. Rather than entering legacy institutions with Silicon Valley hubris, their strategy relies on Forward Deployed Engineers (FDEs) who embed directly inside public agencies to co-own operational problems. Peregrine treats forward deployment not as a customer support cost center, but as its primary engine for R&D and product discovery. Solutions engineered in the field regularly evolve into core platform primitives or specialized tools. For example, direct exposure to customer workflows led engineers to develop internal property-editing pipelines, an emergency hurricane simulator, and a municipal fire department resource-placement calculator that models response coverage against city budget constraints. Autonomous Integration and Long-Horizon Case Reasoning (27:08) To handle tens of thousands of disparate public datasets across multiple jurisdictions, roughly 50% of Peregrine's engineering team focuses on platform architecture. The company utilizes background AI agents that autonomously write approximately 90% of the underlying integration code in Python notebooks under human review, leveraging deterministic evaluation benchmarks to maintain accuracy. For operational agencies, Peregrine applies reasoning agents to unstructured, multimodal records. In a Florida county facing more than 100 water rescues in a month, the system synthesized 911 logs, historical weather data, and terrain maps to identify a three-day weather pattern that formed hazardous underwater sand channels and rip currents. In criminal investigations, Peregrine developed a 'Cold Case Agent' capable of autonomously analyzing 200 to 300 gigabytes of mixed evidence over 30 to 60 minutes. The agent validated its reasoning by reproducing the evidentiary findings of a past wrongful-conviction exoneration and later helped Wisconsin investigators place a suspect at a remote disposal site by isolating cell tower pings hidden within call logs. Quiet Professionalism and Scaling to Local Agencies (34:25) After navigating hiring headwinds during the 2020 social unrest and debates around police funding, Peregrine maintained an apolitical posture and expanded its focus beyond police departments to fire, emergency medical services, disaster response, and county health agencies. The founders maintain a 'quiet professional' ethos, deliberately refusing to claim public credit for criminal investigations or market case successes in press releases, which they argue risks eroding institutional trust with public servants. This deference extends to morally complex capabilities like facial recognition, where Noone rejects vendor-imposed rules in favor of helping local governments understand legal boundaries and local community standards. By vertically integrating its entire software architecture—from ETL pipelines and ontologies to APIs—Peregrine reduced marginal deployment costs well below $1,000,000 annually. This cost reduction allows the company to move past eight-figure enterprise contracts and target an operating model capable of supporting 10,000 smaller city and county governments. Peregrine's expansion model relies on making advanced municipal software affordable to smaller local agencies while adhering strictly to local governance and privacy laws. However, scaling infrastructure across thousands of public institutions requires continuously navigating polarized public attitudes toward law enforcement, high public distrust of AI, and varied local regulations regarding sensitive municipal data.

Peregrine联合创始人Nick Noone与Ben Rudolph提出，公共安全是城市繁荣的基石，但科技企业不应通过铺设硬件传感器构建中央集权式的监控网络；该公司主张以“反网络效应”与本地数据主权为核心，通过软件集成治理市政既有数据，并依托长流程自主AI智能体与驻场工程服务提升地方机构的决策效能。 颠覆传统监控模式与捍卫数据主权 (18:16) 在公共安全科技领域，传统头部厂商通常依赖售卖专有硬件传感器（如车牌识别摄像头、执法记录仪和监控探头）来捕获并囤积增量数据，借此锁定客户持续增购设备。Nick Noone指出，Peregrine采取了截然相反的路径，实行“零传感器、零新数据采集”模式。该平台定位为纯软件集成与治理层，专注于打通、梳理和赋能地方政府已经拥有但难以调用的存量异构数据系统，从源头上消除了扩张监控国家的商业激励。 针对外界对中央集权式监控系统的担忧，Ben Rudolph强调Peregrine采用“反网络效应”架构，坚持严格的细粒度权限控制与本地数据主权。所有数据资产的绝对所有权均归属于地方机构及社区，平台绝不进行跨机构的数据中央汇聚。同时，在涉及人脸识别等高争议技术时，公司拒绝以硅谷视角向地方政府强加规则，而是尊重全美各城市不同的法治与社区民意标准，依托合规软件取代过去“用汽车后备箱运送物理硬盘跨区共享数据”的低效且失控的高风险做法。 摒弃傲慢的驻场工程与双向研发闭环 (1:28) Peregrine的创立融合了两位创始人截然不同的早期经历：Noone曾领导Palantir的特种作战司令部（SOCOM）技术单元深入中东前线，而Rudolph则在联合国难民署（UNHCR）哥伦比亚及苏丹边境开展人道主义救援，并在Dimagi参与印度国家级结核病服药依从性追踪项目。Rudolph在此过程中确认了关键假设：许多重大现实危机本质上是底层数据管理失序的下游产物。2018年初，在连续遭到20多个机构拒绝后，他们以极度的谦逊说服北加州圣帕布洛警察局（San Pablo PD）指挥官Brian Bubar，获得了进驻办公的许可，从而正式启动实际业务。 Noone认为，硅谷技术人员进入洛杉矶警察局等传统体制机构时常伴随不恰当的优越感，忽视了从业者数十年的复杂业务语境。Peregrine提倡“前线部署工程师”（FDE）机制，要求工程师深入现场共担问题，并以“静默的专业人士”自居，不通过张扬个案战果来做自我公关。FDE团队并非传统的售后技术支持，而是核心研发的驱动引擎：早期工程师在现场快速构建的补丁被沉淀为平台通用编辑功能，针对特定机构开发的飓风模拟器与消防站选址预算计算器等工具，也持续反哺着整个产品架构。 自主智能体驱动的数据治理与深度侦查 (22:57) 在平台工程层面，Peregrine将约50%的核心工程人力投入到基础数据底座的开发中，约95%的后端精力集中在查询前的数据准备与语境化构建上。为处理全球数百个管辖区数万个存量公共部门数据集的对接难题，公司开发了后台长期运行的集成AI智能体。在人类部署策略师的监督下，这套多智能体编排系统能够自主解析遗留数据库模式、推断领域本体并在Python环境中编写执行脚本，平台底层约90%的集成代码均由此类自动化智能体编写。 在分析终端，语义嵌入与推理智能体推动公共安全机构从基础关键字检索迈向深度因果挖掘。例如在佛罗里达某县单月超100起水上救援案例中，智能体综合911调度、历史气象与地形数据，成功识别出为期3天的特定天气周期冲刷水下沙槽导致致命离岸流的深层规律；在犹太教堂威胁调查中，语义模型超越精准关键词限制，成功串联起跨区域分散的反犹隐蔽威胁。针对谋杀重案积压问题，Peregrine的“悬案智能体”（Cold Case Agent）可在30至60分钟内自主解析200至300GB的多模态卷宗（包含监控音视频、纸质扫描件与通话记录），在威斯康星州的一起凶杀案中，该系统精准提取出深埋于海量通讯记录中的关键基站信号，迅速锁定了犯罪嫌疑人在现场及抛尸地点的行踪轨迹。 超越政治极化并降本赋能中小城市 (14:19) Peregrine在2010年代末于旧金山起步时，适逢全美围绕“撤资警察”（Defund the Police）的激烈社会撕裂期，技术人才招聘曾面临严峻阻力。创始人团队坚持非政治化的高同理心原则，每日往返于湾区抗议活动与警局地下室凶杀重案组之间。随着时间推移，Peregrine将业务边界稳步拓展为覆盖警察、消防、急救（EMS）、应急灾害管理和县级公共卫生部门的综合市政平台，通过全栈垂直整合自研的ETL数据管道、治理权限体系与UI层，彻底摊薄部署成本。 Noone强调，传统企业级软件服务地方安全机构失败的主因在于单点部署边际成本过高，往往只聚焦数千万美元级的头部大单。Peregrine通过现代软件抽象将年均部署成本压缩至远低于100万美元，使其能够以平价方案向各类下沉的中小县市及被忽视的基层机构推广，并设定了未来为全球一万座城市提供基础设施记忆层的长期发展规划。 Peregrine展示了一种区别于传统硅谷硬件扩张范式的公共安全技术路径，但其赋能一万座城市的远景规划，仍高度依赖其能否在严格遵循地方隐私法规和数据主权的前提下，持续消化不同层级政府极其繁复的陈旧遗留系统，并在低廉成本约束与重度定制化信任部署之间维持长久平衡。

## Background 本期对谈由红杉资本（Sequoia Capital）于 YouTube 平台发布，由红杉合伙人 Sonya Huang 主持。视频围绕 Peregrine 联合创始人 Nick Noone 与 Ben Rudolph 的创业历程展开，深入探讨了公司如何颠覆传统公共安全科技依赖硬件传感器的采集模式，在不新增数据、杜绝监控扩张的前提下打通城市既有数据孤岛，并运用长程 AI Agent 协助全美 400 多座城市的警务、消防与应急部门解决积案排查、灾害归因等复杂现实难题。 ## TL;DR Peregrine 颠覆了传统安防企业依靠铺设硬件传感器扩张的逻辑，坚持不收集新数据、坚守数据主权与细粒度权限控制，专注于为城市公共机构打通既有的碎片化数据孤岛。创始团队结合前线部署工程（FDE）模式与深度的现场同理心，将绝大部分工程算力倾注于底层数据治理与清洗，并利用长程 Agent 自动编写 90% 的集成代码，大幅降低了公共机构的技术使用门槛。在应用端，平台演化出能够深入梳理数百 GB 杂乱证据链的“悬案 Agent”（Cold Case Agent）以及归因极端灾害与安全威胁的分析能力。Peregrine 恪守“反网络效应”与静默专业主义，不以技术权威自居干涉本地决策，旨在以垂直整合的技术底座赋能全球上万座城市，守住公共安全的基石。 ## Atmosphere & style 整场对话沉稳、克制且极具现场质感，彻底剥离了硅谷科技圈常见的浮躁叙事与说教姿态。Sonya Huang 的提问直击敏感伦理与商业本质，而两位创始人则以质朴且极具敬畏心的语言，真诚还原了坐在晃动的副驾驶座敲代码、随警通宵巡逻、在社会舆论撕裂中直面两难抉择等鲜活细节。聆听体验既有长程 AI 与数据管线等硬核技术推演的理智严谨，又贯穿着对基层执法与救援人员的深切共情，散发出一种务实、谦卑而充满使命感的“技术管家”气质。 ## Bulletin - [18:16](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1096s) 与传统依赖铺设传感器硬件来扩大数据规模的安防企业相反，Peregrine 坚持零传感器、不收集新数据，专注于打通城市既有数据孤岛并确保数据主权归属当地机构。 - [30:00](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1800s) Peregrine 构建的后台长程 Agent 架构已能自主编写高达 90% 的异构数据集成代码，支持自动化解析本体结构并生成 Python 脚本。 - [33:04](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1984s) 其研发的旗舰“悬案 Agent”成功在 300GB 卷宗与海量基站详单中锁定命案嫌疑人关键轨迹，并独立复现了人工侦探对冤案的脱罪推演结论。 - [43:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2625s) 团队将前线部署工程师（FDE）视为研发与增长引擎而非成本中心，借助 AI 将定制化服务成本大幅压缩，目标是成为全球 10,000 座城市的机构记忆底座。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=0s) 城市繁荣与社会信任的基石是公共安全，兼具客观物理安全与主观心理安全感，现代数据基础设施必须在捍卫隐私的前提下保障这一底座。 - [02:07](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=127s) 前线部署工程（FDE）的核心是彻底认领客户业务结果并以数倍速度交付，科技精英需摒弃自命不凡的傲慢，以同理心沉入机构土壤。 - [05:23](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=323s) 难民署和偏远地区医疗援助的实战经验证明，现实世界中绝大多数错综复杂的系统性困境，其本质都是“数据问题”的下游副产物。 - [10:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=645s) 在遭遇二十多次拒绝后，Peregrine 通过深入研读警员破获的“红河行动”案情并以谦虚求教的姿态，敲开了圣巴勃罗警局的第一扇大门。 - [14:19](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=859s) 在“削减警费”与社会两极分化的风暴中创业，创始人通过亲赴一线建立灰色认知，坚持构建脱离政治意识形态的公正技术工具。 - [18:16](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1096s) 彻底颠覆传统数据采集模式：不添加传感器，不收集外部新数据，而是将分散在既有孤岛中的数据以高精度、强治理方式激活。 - [21:20](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1280s) 奉行去中心化的数据主权哲学，数据 100% 归属于当地机构，通过严格的细粒度数字围栏和审计轨迹打破“全景敞视式监控国家”隐忧。 - [22:50](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1370s) AI 应用从简单的卡片搜索走向深度归因分析，成功在佛州通过气象、地形与报警记录推断出离岸流沙道成因，并跨辖区串并犹太会堂威胁。 - [26:34](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1594s) “95% 的苦工发生在提问敲下之前”，团队绝大多数工程资源投入在底层数据对齐与可溯源验证，借助大模型代码生成为前线快速构建应急模拟器。 - [30:00](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1800s) 长程后台 Agent 通过主调度与子 Agent 协同接管了 90% 的数据集成代码编写，结合确定性评测集消除了人工接入的劳动密集型瓶颈。 - [33:04](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1984s) 针对 200 至 300 GB 案卷的“悬案 Agent”可在 30 至 60 分钟内完成深度时空关联分析，免去刑警数百小时的人工音视频排查苦役。 - [34:25](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2065s) 坚持“静默的专业主义”与“反网络效应”，把聚光灯留给一线公职人员，绝不通过掠夺性汇集数据谋求集权式的平台网络效应。 - [38:40](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2320s) 在人脸识别等伦理抉择上遵循“遵循客户与法律”准则，拒绝以硅谷导师姿态划定道德红线，专注帮助地方机构厘清政策边界与社区共识。 - [40:48](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2448s) 为预算有限的“弱势机构”打造全自研、深度垂直整合的技术栈，不靠应付式的拼凑组件，以极高标准建立底层护城河。 - [43:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2625s) 前线部署工程师（FDE）模式是业务扩张与产品研发的核心飞轮，一线工程师现场拼凑的应急方案往往反哺出平台通用的核心产品基元。 - [48:50](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2930s) 终局目标是成为全球 10,000 座城市的机构记忆底座，通过工程手段降低边际交付成本，对公共治理权力保持敬畏，保护城市多样性。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=0s) 缘起与愿景：用数据基础设施让城市更美好 Peregrine 的诞生植根于一个清晰而宏大的愿景：如何利用前沿科技赋能城市、县区及州级政府，深刻改变人们日常生活的真实空间。在构建“让城市变得更美好”（Making Cities Awesome）的蓝图中，底层的绝对基石是“安全”（Safety）。这种安全具有双重属性：既需要客观上的现实安全，也需要公众心理上切切实实的“安全感”。当这种稳定作为底座被确立时，城市社会所能释放的潜力和可能性是不可限量的。 在创业构想之初，团队反复推演并试图平衡一系列看似对立但内在紧密相关的命题。一方面，他们希望将技术系统深入部署到掌握海量极端敏感数据的庞大公共部门内部，彻底杜绝数据在流转与处理过程中的误用和滥用；另一方面，他们致力于打造现代城市的核心骨干数据基础设施，在坚守与捍卫个人隐私、个体独立性（Individuality）的同时，搭建起促进社会融合的桥梁。只要人们能在安全信任的环境中共享同一种真实透明的信息视野，大家其实都会朝向一致的公共福祉迈进。 ```mermaid flowchart TD accTitle: Peregrine's Urban Safety and Data Infrastructure Vision accDescr: A structural hierarchy showing how modern infrastructure leads to individual privacy, safety, and community thriving. subgraph Impact["城市繁荣与社会凝聚"] Thriving["社会繁盛与群体互信"] end subgraph CoreNeed["需求底层基石"] ObjSafety["客观安全 (Objective Safety)"] SubjSafety["主观安全感 (Feeling Safe)"] end subgraph Foundation["底层数据基础设施"] Backbone["现代城市数据基础设施 (Peregrine)"] PrivacyGov["杜绝数据误用 / 捍卫个体隐私"] end PrivacyGov --> Backbone Backbone --> ObjSafety Backbone --> SubjSafety ObjSafety --> Thriving SubjSafety --> Thriving ``` ### [02:07](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=127s) 前线部署工程师（FDE）的本质与硅谷的认知盲区 主持人 Sonya Huang 提及 Nick Noone 曾掌管 Palantir 的特种作战司令部（SOCOM）业务单元，并深入中东等危险且高风险的情报行动一线，借此切入核心问题：究竟什么是前线部署工程（Forward Deployed Engineering, FDE）？硅谷对这种模式究竟存在哪些误解？Nick 坦言，在加入 Palantir 之前他对这个概念毫无认知，只看到身边的同事不断奔赴客户一线现场“做着神鬼莫测的硬核任务”，然后回办公室分享惊心动魄的故事。对他而言，深入客户的实际业务语境、把客户的问题当成自己的问题去全权负责，是一场梦寐以求的冒险。 在 Nick 看来，真正的 FDE 文化建立在两个深刻的共识之上： - **全权共担结果（Outcome-Driven）**：工程师心理上必须完全认领或共同承担客户的问题。团队内部常强调要“与客户并肩、代表客户一路走到终局”，技术交付只是手段，核心是以比任何普通团队快 3 到 5 倍的速度，不惜一切代价为客户拿下最终结果。 - **胜利永远属于客户**：必须在文化上清醒认知，最终的军功章属于一线客户，而非外部工程师团队。以跨越桌案、对客户机构和一线人员产生的切身现实影响为荣，才是团队力量的源泉。 针对“硅谷做错了什么”，Nick 指出，顶级名校背景（如全美前五、前十院校）的顶尖技术人才往往习惯带着精英傲气走进像洛杉矶警察局（LAPD）这样复杂的公共机构。这些机构拥有深耕数十年的制度惯性、人事生态与协同决策机制。硅谷人习惯把“快速推进”强加给对方，这种姿态极易被一线人员解读为傲慢与冒犯，因为一线从业者往往为了这项职责奉献了整整一生。技术精英必须放下自视甚高的智商优越感，搁置个人光环与傲慢，以极致的同理心（Empathy）与长期耐受度沉入客户土壤，方能做成事情。 ### [05:23](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=323s) 难民署与基层医疗：下游问题皆为数据问题 与 Nick 的军事情报背景形成强烈对照的是联合创始人 Ben Rudolph 的早期经历。毕业时，Ben 面临两个走向极端的分岔路口：一个是当时初露锋芒的独角兽 Airbnb，另一个则是联合国难民署（UNHCR）。出于对“如何用技术产生真实世界影响”的执着，Ben 毅然选择了难民署，奔赴苏丹边境以及哥伦比亚边境，协助那些因战火流离失所、冒着生命危险跨越国界的难民群体。 在极其严峻的极端地带，Ben 深刻体悟到技术的能与不能：联合国难民署做着极其崇高的工作，但组织内部处于极度割裂的状态，数据往往分散孤立在简单的电子表格（Spreadsheets）中，混乱而难以汇总利用。这次经历让他顿悟了一个核心论点：**现实世界中的诸多复杂系统性困境，本质上都是“数据问题”的下游副产物（Downstream of data problems）**。 离开难民署后，Ben 加入了专注于全球资源匮乏地区最后一公里医疗解决方案的初创公司 Dimagi。在印度，他深度参与了印度政府的一项全国性项目，通过开发轻量化移动应用，协助偏远农村地区的结核病（Tuberculosis）患者按时服药遵医。这项技术在印度全国范围内迅速铺开，挽救了海量生命。正是这些跨越人道救援与公共卫生的实战经历，构成了 Peregrine 核心理念的另一半拼图：将强大的数据基座赋能给最需要现代科技支撑的本地社区、城市、县域与州政府。 ### [08:25](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=505s) 为什么是城市？为什么是公共安全？ 2016 年，在毕业近十年后的一场大学体操队队友度假聚会上，Nick 和 Ben 重新聚首并深谈，决定共同创立一番事业。彼时 Nick 长期往返于巴格达与中东险境，Ben 则在非洲与印度各地常年穿梭。在旧金山坐下来反思过往经历时，他们意识到：全球的宏大叙事固然重要，但城市与社区才是普通人切切实实生活、成长并度过绝大部分生命时光的地方。 他们解构了美国现代城市的运转逻辑，发现城市繁荣金字塔的最底座始终是“公共安全”。若没有基础的客观安全与心理安全感，任何教育、商业与文化繁荣都无从谈起。当时，他们面临的技术与治理挑战异常具体且富有张力： - 如何在进入高度复杂的公共机构深处部署系统的同时，从架构上彻底根绝其内部敏感数据的泄露与不当使用； - 如何在打造现代城市数字骨干的同时，恪守个体隐私保护的红线； - 如何通过打破数据孤岛，让持不同立场的人在面对同一客观事实时凝聚共识，打破彼此的猜忌与隔阂。 ### [10:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=645s) 踏破铁门：连续被拒两打后敲开圣巴勃罗警局 在将理念付诸实践的过程中，Peregrine 经历了极为艰难的破冰期。在最初尝试接入真实的城市公共安全部门时，他们遭遇了来自各个机构冷酷且连续的拒绝，Nick 回忆被拒绝的次数“绝对超过了两打（24 次以上）”。 转机出现在 2018 年初： - **正式进驻日期**：2018 年 2 月 26 日（公司在约一个月前完成注册成立）。 - **破冰之地**：位于北加州的圣巴勃罗警察局（San Pablo Police Department）。该部门正式给他们签发了工牌，腾出了办公桌位，并赋予了他们直接接触核心业务数据的权限，允许工程师团队在警局内部扎根。 - **关键推手**：团队并非盲目推销，而是在前期系统性调研了全美在市政级公共安全创新领域的领军人物。他们读到了一篇关于圣巴勃罗警局年轻且极具前瞻视野的指挥官 Brian Bubar 的深度报道，以此为线索叩开了这扇至关重要的大门。 ### [11:41](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=701s) 敲开第一扇门：圣巴勃罗警局与“红河行动” 在经历了二十多次碰壁后，Nick Noone 和 Ben Rudolph 终于迎来了转机。当时，圣巴勃罗警局（San Pablo PD）有一位名叫 Brian 的警官，他在警局这种讲究资历与年资的半军事化架构中晋升极快，年轻、聪颖且极具创造力。Nick 和 Ben 在网上读到了一篇关于“红河行动”（Operation Red Reach）的报道，那是当时北加利福尼亚州极具代表性的一起跨司法管辖区的帮派毒品调查大案。Brian 当时作为一名卧底侦查员，正是打破案件僵局的关键人物。这一行动不仅登上了新闻头条，更彻底改变了康特拉科斯塔县（Contra Costa County）的治安局面。 正是这份报道让两位创始人鼓起勇气直接拨通了 Brian 的电话。他们的开场白毫无硅谷创业者的傲慢，而是坦诚与谦逊：“Brian，我们懂的并不多。我们或许能帮上一点忙，但其实也说不准。我们能否过去拜访你，问几个问题，向你请教，了解你们曾经做过的那些了不起的破案工作？也许随着时间推移，我们能一起做点东西，看看到底能带来什么。”正是这种完全以建立信任为先的态度，让 Brian 决定给这群年轻人一次机会。 - **破冰契机**：创始人没有靠强推技术，而是靠深入研究对方破获的代表性案件（Operation Red Reach）并主动示弱请教，才赢得了警方的信任与接待。 - **初次通宵随警巡逻（Ride-Along）**：早在早期调研阶段，团队就参加了奥克兰警局（Oakland PD）的通宵巡逻。当时硅谷去参加巡逻的通常是记者或社区公益人士，极少有工程师。这次经历让他们切身体会到执法机构极难渗透，也让他们意识到必须放平心态、慢慢扎根。 --- ### [14:19](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=859s) 在“削减警费”与社会动荡的风暴中创业 Peregrine 创立于 2017 至 2018 年前后，不久便遭遇了 2020 年新冠疫情与“削减警费”（Defund the Police）浪潮的双重冲击。那段时期美国社会撕裂严重，作为一家专注服务警务部门的初创公司，处境极其微妙与艰难。Nick 坦言自己是一个偏好体验式、触感式学习的人，习惯通过实际行动自下而上建立对事物的认知，这种思维方式促使他时刻质疑既有假设，但并非出于盲目犬儒，而是渴望探寻真相与保持好奇冒险精神。 这种探索让创始人们每天都在承受巨大的心理张力：一方面，他们大学同窗、曾经并肩奋斗的好友们站在旧金山妇女中心外举牌抗议警察；另一方面，他们每天清晨开着一辆本田雅阁（Honda Accord）穿过海湾大桥，进入对岸警局与重案组、帮派命案侦查员并肩工作。Nick 认为，坐在外面评头论足毫无意义，真正的特权是“亲自跳进泳池去游泳”。唯有深入现场，才能发现事情远非非黑即白，其中充满了复杂的灰色地带与人性纹理。这种两极分化甚至波及了招聘——当时他们给工程师打电话甚至直接被拒接。但他们坚信，只要能抛开意识形态，为基层机构和市民提供真正需要且能解决实际问题的非政治化工具，这项事业就具备深远的价值。 ```mermaid flowchart TD accTitle: 动荡时期的双重现实与同理心建立 accDescr: 创始人通过深入一线，在两种冲突的社会现实中建立对复杂问题的理解。 subgraph Tension ["社会两极分化与认知撕裂"] A["旧金山湾区同窗好友<br/>（街头抗议警方）"] B["对岸警局一线侦查员<br/>（调查恶性帮派命案）"] end subgraph Founders ["Peregrine 创始人的亲身实践"] C["每日开本田雅阁往返海湾大桥<br/>身处两种现实交界处"] D["拒绝局外审判<br/>跳入池中建立深度同理心"] end subgraph Outcome ["核心产品理念定位"] E["超越政治化立场<br/>提供客观、安全、公正的技术支撑"] end A --- C B --- C C --> D D --> E ``` - **从单一到多元**：Peregrine 早期仅服务于警务部门，而今已拓展至消防救援、应急管理、公共卫生等整个城市公共安全生态。 - **招聘困境**：在“削减警费”呼声最高时，许多硅谷工程师听到为警察做软件便直接拒绝甚至不回消息。 --- ### [18:16](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1096s) 颠覆传统模式：不做数据收集，专注孤岛打通 当被问及 Peregrine 与 Flock、Axon 等传统公共安全科技公司的本质区别时，Nick 明确指出：以往的大多数公司都是建立在“数据采集”（Data Collection）的逻辑之上。无论是安装硬件摄像头、传感器被动搜集，还是靠人工系统直接录入，这些企业都是靠不断向客户系统里填充并存储更多数据来扩展版图，进而利用渠道优势兜售更多采集硬件。 Peregrine 则彻底把这个模型反了过来（Inversion of the Collection Model）： 1. **零传感器与不引入外部新数据**：Peregrine 坚决不从事向客户灌输其未曾拥有数据的数据采集生意。 2. **连接孤岛与深层治理**：专注于在客户既有的分散系统之上构建一层高度安全、权限严格且受有效监管的智能层。 3. **提升精度而非扩张数据量**：核心价值不是最大化聚集数据以追求“数据网络效应”，而是协助一线公职人员更精准、更合规地与现存数据交互，打通信息孤岛。 对于外界所担忧的公共或私营部门建立“集中化全知全能机构”（Authoritarian Big Brother）的监控风险，Peregrine 的策略是用去中心化思路加以消解。客户的根本痛点从来不是数据不够多，而是无法以安全、高信任度的方式将已有数据用起来。 --- ### [21:20](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1280s) 数据主权与权限治理：破解“监控国家”与公共安全的矛盾 Ben 强调，从 Peregrine 创立的第一天起，细粒度权限控制、数据治理（Data Governance）和数据主权（Data Sovereignty）就被写入了产品底层逻辑。在面对这种极高责任风险的政府机构时，这些规范不是锦上添花的特性，而是进驻的前提条件。 解决“公共安全”与“避免成为监控国家”之间内在张力的唯一方法，就是严格捍卫本地智能与数据封闭： | 维度 | 传统数据采集公司 | Peregrine 架构哲学 | |:--- |:--- |:--- | | **数据归属** | 数据汇集进厂商云端平台，往往衍生厂商所有权与二次利用 | 数据百分之百属于当地机构及其服务的市民，Peregrine 绝不据为己有 | | **商业逻辑** | 靠铺设更多传感器扩大数据池，形成强网络效应 | 纯软件层驱动已有系统，专注本地数据协同与精度计算 | | **共享控制** | 往往粗放打通或依赖传统线下传递（如纸质、U盘拷贝） | 提供极其严格、精准的颗粒度控制，仅在必要时安全受控地共享片段 | Ben 特别提到，在缺乏合规的现代软件解决方案前，警方为了跨区域协同办案，往往只能把大量纸质材料或数据硬盘装进警车后备箱，一路拉过整座城市交给对方。这种原始方式既不可控也没有审计轨迹，更容易导致隐私泄露。Peregrine 赋予机构的是清晰界定权限的能力，让公职人员能在严格的数字围栏与审计追踪下完成协作，从而真正兑现对公众与社区的信任承诺。 ### [22:50](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1370s) 从“好看的搜索”到深度归因分析：离岸流与犹太会堂威胁调查 当人工智能技术最初部署到公共安全与政府应急机构时，最常见的初步应用往往只是 Nick Noone 和 Ben Rudolph 所称的“好看的搜索”（nice search）——比如搜寻某个特定地址时，警员或分析师不再需要逐行比对繁琐的数据库表格，而是能看到以排版整洁的卡片形式呈现的历史事件记录。然而，随着一线用户对系统的熟悉程度加深，整个机构的应用下限被彻底抬高，开始涌现出一批此前完全无法想象的深度分析案例。 其中一个极具代表性的场景发生在佛罗里达州的一个县级机构。当地在一个月内执行了超过 100 次“水上救援”（water rescue，即在洪涝灾害中救援被困于熄火车辆或淹没房屋中的民众），当地应急部门对此大惑不解，因为历史上从未发生过如此高密度的救援需求。于是他们开始在 Peregrine 系统中深入探究原因。AI Agent 在进行多轮推理与深度检索后，从海量历史事件、911 报警记录和气象数据中抽丝剥茧，发现了一种此前从未出现过的极端规律：某种特定的气象模式罕见地连续持续了整整三天，而这种气象在海底冲刷出了沙质暗道（sand channels），为致命的离岸流（rip currents）创造了完美温床，最终导致大量在海水中受困的险情。这种将事件信息、911 呼叫中心数据、气象记录与地理地形图层融会贯通的能力，直接为一线机构提供了可立即采取行动的决策依据。创始人感慨道，如果团队只是坐在硅谷总部凭空构想飓风救援人员的需求，是绝对不可能预料到这种真实场景的。 另一个典型案例来自一名调查针对犹太会堂（synagogue）安全威胁的侦探。该侦探希望查明该地区两所犹太会堂此前是否还发生过其他类似的反犹太主义威胁事件。在传统关系型数据库中，侦探极难构思出能精确覆盖所有隐蔽或口语化表述的关键词；但借助底层的语义搜索（semantic search）与嵌入向量技术（embeddings），AI 能够从海量非结构化文本中精准理解深层语义，抽取出隐藏在各种散碎记录中的关联模式，成功定位了针对这两座宗教场所的连环威胁网络。 ```mermaid flowchart TD accTitle: 复杂公共安全场景下的多源数据整合与 AI 分析流程 accDescr: 展示了 Peregrine 如何将多源异构数据与地理地形结合，通过语义检索与长程推理输出高价值行动建议。 subgraph Inputs ["多源异构数据接入"] A["911 报警电话与警情记录"] B["历史气象与灾害数据"] C["非结构化自由文本 / 笔录"] end subgraph Processing ["Peregrine 数据融合与 AI 推理"] D["语义嵌入 (Embeddings)"] E["因果推断与模式挖掘 Agent"] F["地理空间地形图层映射"] end subgraph Outcomes ["可落地的业务成果"] G["自然灾害归因：离岸流与沙道成因"] H["案件关联：跨会堂反犹威胁语义串并"] end Inputs --> Processing D --> E E --> Outcomes F --> Outcomes ``` ### [26:34](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1594s) 95% 的苦工在提问之前：低成本软件生成与前沿部署团队 Sonya Huang 指出，正是因为 Peregrine 在后台把所有上下文串联、打通，前端的 AI 才能顺理成章地大显身手。对此，Ben Rudolph 直言：“95% 的工作其实都发生在一线用户敲下问题之前。”如何让 AI 能够精准无误地回答问题并精准溯源（cite）引用的证据，是耗费团队大量工程点数的硬核难题。与外界对“大模型即插即用”的幻想不同，Peregrine 将绝大部分研发精力都投入到底层的数据准备和环境构建上，以确保 AI 面向复杂业务时的准确性与合规性。 伴随着大模型带来的代码生成成本断崖式下跌，Peregrine 内部涌现出一种全新的工作范式：为前沿部署团队（deployment strategists）打造一套集成了严格数据安全、治理控制以及标准 API 的底层平台，使他们能够以极低的成本为客户“用软件重构世界”。并非所有客户都需要一个通用的智能体，某些机构可能急需的是一个定制化的飓风模拟器（hurricane simulator）。通过 AI 辅助编码，部署团队能够在几小时或几天内为特定客户快速交付可运行的应用原型。 关于未来客户是否会完全绕过技术支持团队自主构建应用，Peregrine 表示平台已经 100% 支持这种端到端操作，但目前最主流的模式依然是部署团队作为技术顾问与一线客户紧密合作。Ben 强调，技术的形态始终是“涌现的”（emergent），Peregrine 最核心的护城河是对执法部门、消防部门和急救人员业务痛点的深刻理解与同理心；无论是做图表分析还是现在的 Agent，只要底层的原始数据质量不过关，前端生成的一切分析就都是无用的废纸，因此核心战略始终坚如磐石。 - **核心工程分配**：团队绝大多数工程资源倾斜在数据清洗、预处理、对齐与溯源验证上，而非单纯调用前端大模型 API。 - **敏捷交付形态**：部署团队根据客户特定危机需求（如飓风路径推演、特殊资源调配）利用内部低代码/AI 辅助工具敏捷构建定制软件。 - **技术涌现本质**：前端展示层（图表、报表或 Agent）随时可变，决定成败的关键永远是高保真度的数据基础与业务场景认知。 ### [30:00](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1800s) 长程 Agent 自动编写 90% 数据集成代码 在涉及后台长程运行的 Agent（long-horizon background agents）方面，Peregrine 展现出了工程落地的硬核实力。公司约 50% 的核心研发工程师专注于构建底层数据平台，该平台的目标是协助部署团队将客户自有的异构数据进行系统化接入。截至目前，Peregrine 已经跨数百个机构累计完成了数万个数据集的融合。为了消除这种劳动密集型瓶颈，团队专门构建了一套用于自动化集成的 Agent 架构。 目前，Peregrine 内部采用定制版 Python Notebooks 来编写数据集成脚本，其中高达 90% 的集成代码已经完全由后台 Agent 自动生成，部署团队仅负责高阶逻辑审查。这套 Agent 能够自主在后台运行数个小时：主调度 Agent（orchestrator agent）负责检视客户的原始数据库、推断数据本体结构（ontology）、将庞大的集成任务拆解分配给多个子 Agent（sub-agents），各子 Agent 分头执行具体的连接与转换逻辑并汇总反馈。 ```mermaid flowchart LR accTitle: 长程数据集成 Agent 自动化架构 accDescr: 展示主调度 Agent 接收原始数据后，拆解任务给子 Agent 并基于确定性评测集输出 Python 脚本。 A["客户私有异构数据库"] --> B["主调度 Agent (Orchestrator)"] subgraph Execution ["并行长程任务 (数小时)"] B --> C1["子 Agent A: 本体与结构解析"] B --> C2["子 Agent B: 字段映射与清洗"] B --> C3["子 Agent C: 逻辑校验与代码生成"] end Execution --> D["确定性评估系统 (Eval Set)"] D --> E["集成完成: 90% Agent 编写的 Python Notebooks"] ``` Sonya Huang 评论称，这极其类似于开发者利用 Coding Agents 进行长周期、枯燥无味的底层代码库迁移（codebase migration）。Nick 与 Ben 对此深表赞同，并指出这类工程任务最美妙的地方在于它具有“确定性的可验证性”（verifiable）。Peregrine 为此搭建了极为严苛且客观的集成评测集（eval set），可以在数小时的后台运行后，以数学上的确定性评测 Agent 交付成果的完整性与正确性，从而避免了主观幻觉带来的系统风险。 ### [31:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1905s) 运营端突破：针对几百 GB 卷宗的“悬案 Agent” 在攻克了底层集成的长程任务后，Peregrine 将第二项重兵部署在了面向一线业务结果的专用 Agent 上。团队打造的首个旗舰级业务智能体是“悬案 Agent”（Cold Case Agent）。在许多重大恶性案件或积压已久的悬案调查中，办案人员面对的数据体量极其惊人，往往一次性就需要上传 200 到 300 GB 的海量证据包，其中混杂着执法记录仪视频、监控音频、现场勘查照片以及数以万计的扫描版 PDF 笔录。 Ben 描述了他在各警局实地走访时目睹的震撼场景：大量警局至今仍保存着一箱箱半人高的纸质发黄卷宗，盒子上乱七八糟地贴着几百张旧光盘（CD），侦探们为了梳理时间线和人际关系网，需要徒手翻阅数月甚至数年。 为了验证悬案 Agent 的实战可靠性，Peregrine 采取了一贯坚持的“与客户协同研发”模式。他们与一家刚刚成功纠正一起冤假错案的机构合作——该案中一名无辜男子被错误定罪入狱多年，侦探们通过耗时极长的手工比对最终才实现平反昭雪。Peregrine 的团队向办案机构提议：能否将当年沉重的数百 GB 原始涉案卷宗与电子物证完全灌入这套悬案 Agent 中，测试算法能否独立、自动地推导出与人类侦探完全一致的脱罪证据链。 ### [33:04](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1984s) 悬案 Agent 的实战突破与威斯康星州 300GB 证据链排查 Peregrine 团队在探索智能体（Agent）落地时，提出了一个核心设想：能否让 Agent 复现一线刑警耗时数月才能完成的调查推演与结论？基于这个思路，团队提取了一起复杂案件的全部证据和底座数据，构建了一个能够进行长程推理的智能体。该 Agent 每次需要自主运行 30 到 60 分钟，进行深度的上下文联想与多源数据交叉验证，最终成功复现了经验丰富的刑侦专家依靠纯人工侦查才得出的关键推断。这一成果直接演化成了 Peregrine 的首代“悬案 Agent”（Cold Case Agent），并在全美多地的警务部门展开部署。 在威斯康星州某县的一个真实案例中，调查人员面对着多达 300GB 的海量非结构化与结构化证据。悬案 Agent 在这片数字汪洋中展开排查，从无数杂乱的基站通话详单（CDR，Call Detail Records）与手机基站定位数据（Cell Tower Pings）中，精准锁定了几条极为零碎的基站信号跳变。就是这几行看似不起眼的数据，不仅将嫌疑人死死锁定在案发第一现场，还直接印证了他曾在弃尸地点出现的时空轨迹。Ben 指出，对于基层公职人员而言，这种长程自主 Agent 的最大价值在于为他们“彻底松绑”——不再需要基层警员花费数十甚至上百个小时枯坐在屏幕前，去逐帧查看监控视频或一秒一秒地倾听录音材料。 ```mermaid flowchart TD accTitle: 悬案 Agent 证据处理与推演流程 accDescr: 展示非结构化多源数据输入后，Agent 通过 30 到 60 分钟自主深度分析输出关键案件线索的过程。 subgraph Input ["海量案卷数据输入"] A["多源证据集 (视频/音频/报告)"] B["300GB 原始数据 / 通话详单 (CDR)"] end subgraph Processing ["长程智能体推演 (30-60 分钟)"] C["时空轨迹关联分析"] D["基站定位信号交叉校验"] end subgraph Output ["调查突破与公职人员减负"] E["嫌疑人现场与弃尸点精准定位"] F["免去数百小时人工音视频排查"] end A --> C B --> D C --> E D --> E E --> F ``` ### [34:25](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2065s) “静默的专业主义”与反网络效应主张 Nick 强调，面对这种极具社会影响力的重大案件突破，Peregrine 在商业文化上恪守着一条铁律：绝不把功劳揽在自己身上，绝不大张旗鼓地向外界吹嘘“看我们用技术改变了世界”。在他看来，高调揽功是摧毁客户信任最快的方式；相反，技术厂商应该做“静默的专业人士”（quiet professionals），隐藏在幕后全心全意赋能客户，把聚光灯留给那些在一线履职的公职人员，这也是 Peregrine 能够持续受邀解决下一个难题的立身之本。Sonya 提道，在如今公众对 AI 不信任度极高的社会舆论下，这种真正造福社会的正面案例其实非常珍贵，理应被更多人听到。 谈到保护犹太教堂安全等反恐反威胁案例时，Nick 阐述了 Peregrine 独特的“反网络效应”（Anti-Network-Effect）业务哲学。当前美国社会对一种中心化、威权式的 AI 监控怪物（Centralized Authoritarian Panopticon）充满了合法性焦虑与深层恐惧，这种“中央全景敞视”是深深违背美国去中心化立国精神的。许多初创公司为了快速拼凑功能，往往盲目依赖开源工具，甚至在法律法规、部门操作规程（SOP）和地方条例的灰色地带乱拉乱采数据。Peregrine 则反其道而行之： - **数据神圣性与主权归属**：不搞中央集权式的数据聚合，坚决捍卫每个独立机构对其数据的所有权和独立工作习惯。 - **保护网络效应壁垒**：公权机构之间天然具有强网络效应，技术公司的职责不是打破规则打破隔离，而是严格在合规红线内打造安全的数据治理（Data Governance）与精细权限控制逻辑（Permissioning Logic）。 - **学会“放手”（Letting go）**：克制为了高速增长而掠夺式扩张的冲动，为客户和选民构建高透明度、高完整性的基础设施。 ### [38:40](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2320s) 人脸识别的道德抉择与“遵循客户与法律”准则 针对 Sonya 提出的敏感道德伦理问题——例如对人脸识别（Facial Recognition）的态度，以及在面临棘手道德权衡时的北极星法则——Nick 给出了明确答复：一家硅谷科技公司自上而下向全行业发号施令、强推某种普适性的价值观或技术禁令，在本质上是狂妄且完全错误的。无论是决定使用某种感知算法，还是制定数据留存周期，Peregrine 都坚决不越俎代庖。 在美国的现实环境中，绝大多数公共安全部门及其所在社区自主选择**不**采用人脸识别技术；部分机构立场极为强硬，部分则基于主观偏好做出判断。Peregrine 的职责不是代替客户去划定道德“红线”，而是做到“遵循客户，遵循法律”： - 帮助客户彻底看清技术落地的现实上下文，厘清隐藏的政策限制与社会影响。 - 耐心辅导并协助部门确立最符合当地社区规范的标准操作流程与技术边界。 - 搭建同行交流与信息互通的桥梁：北美各个城市在遇到技术伦理抉择时，最常见的做法是“互相打电话咨询经验”；Peregrine 致力于让信息在城市间透明流动，让决策者能在获得充分信息后做出最明智、最合规的抉择。 ### [40:48](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2448s) 为底层弱势机构造技术：垂直整合与深入交付闭环 Sonya 指出，Peregrine 的服务对象是那些预算紧张的“弱势机构”（Underdogs），这与同赛道另一知名巨头 Palantir 形成鲜明对比——Palantir 往往只接八位数（千万美元级）以上的巨额政府合同。面对“导师是否劝阻过这种重交付模式无法实现商业化规模扩张”的疑问，Nick 和 Ben 表示，选择这条难走的路需要极大的战略自信。 Peregrine 坚信，只要产品真正解决前人未曾攻克的问题，规模效应就必然会在后期显现。为此，他们不惜投入巨大的人力与工程成本，构建了一套全自研的深度垂直整合技术栈： - **全链路覆盖**：贯穿底层网络接入、权限管控与数据治理、ETL 数据清洗与管线调度、本体论建模（Ontology）、前端 UX 到上层开放 API。 - **第一梯队标准**：每一个技术组件都必须做到第一流水平，坚决拒绝应付合规的“勾选式”（check the box）凑合产物。 - **对最终结果完全负责**：Peregrine 的“前线部署战略家”（Deployment Strategists）会深入客户现场，与客户共同承担解决难题的最终责任，死磕每一个业务结果。这种在早期看似极其手工、耗时耗力且不可扩展的深重交付，恰恰反哺出了一套能够适应极端复杂现实、并在此后高效规模化复制的坚固平台底座。谈及这批低调谦逊、具有极高使命感的前线部署团队时，Nick 透露公司内部建立了一种极其严格、追求卓越的高强度反馈文化，两位创始人通过性格测评与层层筛选，将“对极致的追求”刻进了团队的组织基因。 ### [43:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2625s) 信任一线贡献者与“前线部署工程师”（FDE）模式 Nick Noone 与 Ben Rudolph 深入探讨了 Peregrine 内部对于“信任一线独立贡献者（IC）”的组织文化。Nick 指出，想要把前线部署工程师（Forward Deployed Engineer，简称 FDE）的模式真正做透，极度依赖组织层面的信任。他打了一个非常形象的“黑暗洞穴”比喻：团队就像是把一个人派进一个未知的漆黑洞穴中，腰带上挂着各种工具，虽然洞口处系着一根安全绳、站着同伴，一旦呼救可以被拉回来，但本质上，工程师是孤身深入前线地带的。公司对他们的嘱托是：“坚守你们的诚信与原则，拿着这些工具，去创造价值。” 要实现这种规模化，必须建立坚实的组织信任，坚信真正的创新并非诞生于象牙塔，而是发生在组织最遥远的边缘触角。正是遍布全美乃至多个国家的 FDE 团队，在源源不断地向后方的工程、产品和设计团队输送创新的养分。主持人 Sonya Huang 赞叹道，很多科技公司通常把前线部署或驻场工程师视为一种“成本中心（Cost Center）”，但 Peregrine 的视角完全相反。Nick 与 Ben 肯定地表示，在 Peregrine，FDE 团队绝对是纯粹的研发（R&D）和业务增长引擎。在签下客户进入客户现场后，团队会启动内部试点模型，持续冲刺开拓新的用例，公司从“进场（Land）”到“拓展（Expand）”的核心动能，完全是由冲锋在前线的 FDE 驱动的。 ```mermaid flowchart LR accTitle: FDE 双向飞轮机制 accDescr: 展示前线部署工程师如何连接客户现场真实痛点与后方产品研发平台 subgraph Frontline["前线客户现场"] A["一线客户紧急需求 / 痛点"] --> B["前线部署工程师 (FDE)<br/>快速原型与临时方案"] end subgraph ProductCore["后方核心平台"] B -->|"高保真信号输入<br/>(创新与需求反馈)"| C["工程、产品与设计团队"] C -->|"沉淀核心基元 / 抽象新能力"| D["Peregrine 基础平台"] D -.->|"赋能更强工具"| B end ``` ### [45:34](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2734s) 坐在副驾驶敲代码：从一线痛点反哺产品基元 当 Sonya 询问是否通过这种模式发现了全新产品时，Ben 笑着直言“百分之百，甚至根本无法把这两者割裂开来”。他回忆起公司刚起步时的惊险场景：当时他坐在汽车副驾驶上正开过海湾大桥驶向 San Pablo，Nick 在旁边催促说必须立刻把报告做出来，因为当地刑警 Aaron Blazale 有一项极其紧急的案件请求。Ben 只能在晃动的车里拼命敲击键盘写代码。这种迫切的需求绝不是坐在会议室里空谈理论能凭空想出来的。 Nick 将前线部署对产品演进的推动作用归纳为两大类： 1. **底层产品基元（Product Primitives）的沉淀**：很多核心功能最初源于前线工程师的“野路子”发明。例如，早期 Peregrine 系统根本不支持用户编辑特定字段或属性。为了绕过这个限制并解决客户问题，一名 FDE 甚至基于用户在对象上写的“评论（Comments）”做了一套集成机制——让系统抓取评论内容再去更新数据属性，以此曲线实现了编辑功能。核心团队看到后意识到必须原生支持属性编辑。Nick 强调，公司对工程师的指导原则始终是“以达成最终目标为导向，先别管技术是否优雅”，这种看似难以扩展、像黑客拼凑出的临时方案，反而是技术研发最真实、最敏锐的“信号网络”。 2. **前线创新实验室（Innovation Lab）的单点突破**：另一类则是部署策略师（Deployment Strategist）针对特定客户定制开发的独特能力，虽然可能不需要合并回通用底层，但极具惊艳感。例如，有前线人员直接在 Peregrine 平台内手搓了一个“飓风模拟器”；还有一位刚入职仅仅一个月的女性部署策略师，整合了 911 报警呼叫响应时间、市政预算等多源数据，构建了一个可以在城市地图上虚拟选址规划消防局，并立即测算出对周边居民影响覆盖率的模拟工具。 ### [48:50](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2930s) 终局愿景：赋能万座城市，守住技术敬畏 在对谈的最后阶段，Sonya 提出了一个宏大的未来设想：假定十年后一切进展顺利，Peregrine 已经成为了全球 10,000 座城市的“机构记忆层（Institutional Memory Layer）”，这种超越单一政府任期的巨大影响力伴随着巨大的责任，团队该如何应对？ Nick 将其拆解为“终局目标”与“实现路径”两个维度： - **关于终局（10,000 座城市）**：Peregrine 始终将自己定位为基础设施提供商。技术的目标不是越俎代庖，而是提供底层基础设施，赋能这些公共机构按照自身的意愿和法定职责去掌控与运用数据。Nick 动情地表示，世界上每一座城市、每一个辖区都是独一无二的，保护并延续这种独特的多样性（Potpourri）是一件非常美妙的事，公司的使命是“让每一座城市以它们自己的独特方式变得更加棒（Make cities awesome）”。 - **关于路径与敬畏心**：要走到那一步，必须依靠扎实的诚信、道德指南针与核心价值观，且必须体现在日常的行动而非空洞的口号中。Nick 直言，面对每一个新客户，都需要极度的细致与呵护。在五年前，这类公共安全数据服务几乎是不可能普及的——科技巨头历史上有很多试图涉足州、县、市级公共安全的业务部门最终都以失败告终，因为要提供高度定制化的解决方案，成本高得无法让地方政府承担（年均动辄数百万美元以上）。Peregrine 通过工程效率将服务边际成本降了数个数量级，才赢得了进入这个赛道的入场券。面对数据治理与公共安全，Nick 谦卑地总结：“我们何德何能去认为自己掌控了权力？真正的权力与决定权永远在这些公共治理机构手中。” 最后，Sonya 称赞 Peregrine 团队在如此严肃的领域展现出了极高的审慎态度与管家精神（Stewardship），并表示当下世界急需更多像 Peregrine 这样真正利用 AI 改善公众生活、改善社区的切实范例。访谈在对团队的致谢中圆满结束。 ## Notable quotes > [02:07](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=127s) 与客户并肩、代表客户一路走到终局 — Nick Noone > [05:23](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=323s) 现实世界中的诸多复杂系统性困境，本质上都是“数据问题”的下游副产物（Downstream of data problems）。 — Ben Rudolph > [11:41](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=701s) Brian，我们懂的并不多。我们或许能帮上一点忙，但其实也说不准。我们能否过去拜访你，问几个问题，向你请教，了解你们曾经做过的那些了不起的破案工作？也许随着时间推移，我们能一起做点东西，看看到底能带来什么。 — Nick Noone > [14:19](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=859s) 亲自跳进泳池去游泳 — Nick Noone > [26:34](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=1594s) 95% 的工作其实都发生在一线用户敲下问题之前。 — Ben Rudolph > [34:25](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2065s) 静默的专业人士 — Nick Noone > [38:40](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2320s) 遵循客户，遵循法律 — Nick Noone > [43:45](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2625s) 坚守你们的诚信与原则，拿着这些工具，去创造价值。 — Nick Noone > [45:34](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2734s) 以达成最终目标为导向，先别管技术是否优雅 — Nick Noone > [48:50](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2930s) 让每一座城市以它们自己的独特方式变得更加棒 — Nick Noone > [48:50](https://www.youtube.com/watch?v=JWMF5EpP0KE&t=2930s) 我们何德何能去认为自己掌控了权力？真正的权力与决定权永远在这些公共治理机构手中。 — Nick Noone ## Mentioned Peregrine, Nick Noone, Ben Rudolph, Sonya Huang, Sequoia Capital, Palantir, 特种作战司令部 (SOCOM), 洛杉矶警察局 (LAPD), Airbnb, 联合国难民署 (UNHCR), Dimagi, 圣巴勃罗警察局 (San Pablo Police Department), Brian Bubar, 红河行动 (Operation Red Reach), 康特拉科斯塔县 (Contra Costa County), 奥克兰警察局 (Oakland Police Department), Flock Safety, Axon, 悬案智能体 (Cold Case Agent), Aaron Blazale, Python, Jupyter Notebooks ## Video description Most public safety technology companies grow by collecting more data. Peregrine inverted the model: no sensors, no new data, a business built on connecting the data and information cities already own. Co-founders Nick Noone and Ben Rudolph received more than two dozen no's before San Pablo PD let them in the door in February 2018. Today, Peregrine powers law enforcement, emergency medical services, fire and rescue, and other services in more than 400 cities and communities globally. Nick and Ben explain their north star for data sovereignty, and discuss how Peregrine's philosophy and privacy-first approach to data access and ownership preserves individual privacy and cities' sovereignty. They walk through how AI and long-horizon agents are being deployed: a cold case agent that reproduced an exoneration detectives had reached by hand, a Wisconsin county that placed a suspect using cell records buried in 300GB of evidence, identifying threats to a synagogue, root-causing an escalation in weather-related incidents, and more. Hosted by Sonya Huang, Sequoia Capital 00:00 Introduction 02:07 What Forward Deployed Engineering Means 03:58 What Silicon Valley Gets Wrong 05:23 UNHCR, Dimagi And Downstream Data Problems 08:25 Why Cities, Why Safety 10:45 Two Dozen Nos And San Pablo PD 14:19 Building Through Defund The Police 18:16 The Inversion Of The Collection Model 21:20 Data Ownership And Governance 22:57 From Nice Search To Deep Analysis 29:59 Agents Writing The Integrations 31:45 The Cold Case Agent 35:02 The Anti-Network-Effect Proposition 38:40 Facial Recognition And Hard Decisions 40:48 Technology For The Underdogs 42:54 Trusting The Individual Contributor 48:50 Ten Thousand Cities [Source](https://www.youtube.com/watch?v=JWMF5EpP0KE)

[2:07] Nick Noone: 与客户并肩、代表客户一路走到终局 [5:23] Ben Rudolph: 现实世界中的诸多复杂系统性困境，本质上都是“数据问题”的下游副产物（Downstream of data problems）。 [11:41] Nick Noone: Brian，我们懂的并不多。我们或许能帮上一点忙，但其实也说不准。我们能否过去拜访你，问几个问题，向你请教，了解你们曾经做过的那些了不起的破案工作？也许随着时间推移，我们能一起做点东西，看看到底能带来什么。 [14:19] Nick Noone: 亲自跳进泳池去游泳 [26:34] Ben Rudolph: 95% 的工作其实都发生在一线用户敲下问题之前。 [34:25] Nick Noone: 静默的专业人士 [38:40] Nick Noone: 遵循客户，遵循法律 [43:45] Nick Noone: 坚守你们的诚信与原则，拿着这些工具，去创造价值。 [45:34] Nick Noone: 以达成最终目标为导向，先别管技术是否优雅 [48:50] Nick Noone: 让每一座城市以它们自己的独特方式变得更加棒 [48:50] Nick Noone: 我们何德何能去认为自己掌控了权力？真正的权力与决定权永远在这些公共治理机构手中。

[31:45] Our cold case agent... took all of the evidence and data that they had on that case and started to work on an agent that would run for 30 minutes, 60 minutes to start to glean insights, and eventually we got to the place where it could reproduce the results that those detectives had gotten to.

[0:00] Paragrin is really around this idea of [0:03] how can we leverage technology to work [0:06] with our cities, our counties, our [0:08] states uh to to impact the the places [0:12] that we live in. At the bottom of the [0:14] pyramid really when it comes to how to [0:16] make cities awesome is the idea of [0:19] safety. The idea that people need [0:22] objective safety and also need to feel [0:25] safe. And [0:28] when that stability is there, it's [0:30] amazing what's possible. And there were [0:33] all these kind of contrasting but but [0:36] but actually similar concepts that we [0:39] were wrestling with at the time. We were [0:41] thinking about how to deploy inside of [0:42] institutions and also find ways to [0:45] eliminate the mishandling of information [0:47] inside of really complex organizations [0:51] that have access to super super [0:52] sensitive data. And then if we could [0:54] create an the backbone infrastructure [0:59] level of the modern city then [music] we [1:02] could help to preserve individuality. [1:05] We talk about privacy and other aspects [1:06] of what it means to be like an [1:08] individual person but also bring people [1:10] together. [music] [1:19] [music] [1:28] One of the most important stories in AI [1:30] is from a company that you may not have [1:32] heard of, but is almost certainly [1:33] keeping you and your loved ones safe. [1:36] Paragrin is building AI that protects [1:38] cities and communities and they are [1:40] rejecting the surveillance state while [1:42] doing that. And this should be a [1:44] fantastic conversation about one of the [1:46] most important applications of AI, [1:48] forward deployed engineering, civil [1:50] liberties, and much more. So, let's get [1:52] into it. Nick, you ran Palunteer's SOCOM [1:56] unit. You deployed into dangerous high [1:59] stakes intelligence operations in the [2:01] Middle East. Tell me, what does forward [2:03] deployed engineering actually mean? Um, [2:05] what does Silicon Valley get wrong about [2:07] it? [2:07] >> Palanteer was a really fun experience [2:09] and a really formative one because um [2:11] before joining the company, I had no [2:13] idea what this concept of forward [2:15] deployed engineering meant. Um, but I [2:20] was surrounded by people that were [2:22] running into customer environments and [2:27] doing who knows what. [2:29] >> Yeah. [2:30] >> I watched them I watched them leave the [2:32] office. I watched them come back to the [2:34] office and tell stories. Um, [2:37] it just felt to me like the adventure of [2:39] a lifetime to be able to go into a [2:42] customer context and [2:46] own the customer's problem [2:49] but also [2:50] take pride in the fact that the problem [2:53] is belongs to the customer and then take [2:56] pride in getting the customer to the [2:59] win. And so there these two kind of [3:03] these two truths that [3:06] I learned when I was an early forward [3:08] deployed engineer and that we try to [3:10] embrace as we've built our company and [3:13] those are that [3:15] we psychologically go in and own or [3:18] co-own the problem. We talk internally [3:21] about getting all the way to the outcome [3:24] with and on behalf of our customers. [3:26] technology and all of the the skills and [3:29] ways of delivering our tech is part of [3:34] the answer, but the real answer is just [3:35] getting to the outcome at all costs. [3:37] Ideally, three to five times faster than [3:39] any other person or team could do [3:42] and everything flows from there. [3:45] But recognizing that at the end of the [3:47] day, it's the customer's win. It's not [3:49] our win. and taking deep pride in the [3:52] impact that we have on the institution [3:54] and on the people that we're sitting [3:55] across the table from. Like that's [3:57] that's where the magic is culturally. [3:59] >> What does Silicon Valley get wrong about [4:00] it? When we first started building [4:04] Paragrin, [4:06] the forward to plate engineering concept [4:08] had been around for for a while. And we [4:10] got feedback that when really smart [4:14] people from top five, top 10 or [4:17] equivalent equivalent schools would go [4:20] into a really complicated place like the [4:23] LAPD and [4:27] quickly try to understand [4:30] 30 years of institutional context, [4:33] not just at the data level, but at the [4:35] the human level. Like how does this [4:36] organization work? How are decisions [4:38] made? How do teams interoperate with [4:39] each other? The idea that moving fast is [4:43] even possible in that environment is [4:46] very counterintuitive to the customer. [4:48] And unfortunately, I think that shows up [4:49] as ego [4:50] >> to a lot of organizations where people [4:54] have dedicated their entire lives to a [4:56] particular role. And I think Silicon [4:58] Valley misses that sometimes that [5:02] letting go of our intelligence, letting [5:07] go of our own skills and abilities, [5:09] trying to suspend our ego and then [5:14] really get into the customer's context [5:17] is a is an easy thing to say, but it's [5:19] like it's such a high empathy, patient [5:22] >> Yeah. [5:22] >> way of working. [5:23] >> Yeah. I love that. Ben, you well while [5:27] while Nick was off deploying into the [5:29] Middle East, you were in almost the [5:31] opposite part of the world in every [5:33] sense of the world word. Uh you were [5:35] doing refugee work in Africa, uh in [5:38] India, but you landed on a similar [5:41] thesis of the world and technologies [5:44] role in it. Maybe say a word on that. [5:46] >> Yeah, absolutely. So I for for me I've [5:49] always been obsessed with this idea of [5:53] how do you how do you make impact with [5:55] technology? I graduated school and I had [5:59] these I remember I had these two offers. [6:00] One was to go uh work at a small startup [6:03] at the time, Airbnb, and uh the UN [6:07] Refugee Agency, which is an organization [6:09] that goes and helps refugees uh who are [6:13] making these extremely dangerous uh [6:15] treks across country lines, usually [6:18] because of war, uh to try and get help. [6:20] And so I took that path and deployed to [6:26] the Sudanese border uh the Colombian [6:29] border and you know was deeply humbled [6:33] and uh really un deeply understood uh [6:38] these types of uh situations and where [6:42] where technology can help and where [6:43] where honestly it can't help. And I [6:47] remember the thing that I I took away [6:50] from that experience is, you know, UNHCR [6:52] is this is this deeply disconnected [6:56] organization. It does phenomenal work, [6:57] but it's deeply disconnected. They have [6:59] some data. It's often in spreadsheets, [7:01] and it was very difficult to actually [7:03] make make sense of. And so that's where [7:05] I started to like form this thesis of a [7:08] lot of these problems are downstream of [7:11] data problems where you can have a lot [7:13] of of impact. After UNHCR, the UN [7:16] refugee agency, I left to to join uh an [7:20] organization that you probably haven't [7:23] heard of. It's called Demagi. They're a [7:25] small company. They work building last [7:28] mile healthcare solutions in [7:30] underresourced places around the world. [7:31] And so like one specific example, one [7:33] project I worked on while I was there, [7:35] we were working with the Indian [7:36] government on this project building [7:39] applications to help folks in rural [7:43] India adhere to their uh tuberculosis [7:47] drugs. And uh at the end of the day, we [7:50] did get it deployed and that technology [7:54] rapidly spread throughout the country. [7:56] And that was a really cool moment for me [7:59] and really formed my thesis and idea [8:03] around how Paragan can have impact and [8:05] how you can how technology can have [8:07] impact in uh in the communities that [8:09] that we live in. [8:11] And so paragrin is really around this [8:14] idea of how can we leverage technology [8:18] to work with our cities, our counties, [8:20] our states uh to to impact the the [8:24] places that we live in. [8:25] >> Take us back to 2017. [8:28] You cold called your way into the San [8:30] Pablo Police Department. [8:32] >> What was going through your heads when [8:33] you did that? In 2016 when Ben and I um [8:38] got together actually on a on a vacation [8:40] with a bunch of our gymnastics [8:41] teammates, we um had a conversation [8:45] basically 10 years after we graduated [8:48] saying should we build something [8:50] together? We got together in San [8:51] Francisco and with my experience in the [8:55] national security space and similar in [8:58] the US government, Ben's experience [9:00] working on humanitarian causes. And you [9:03] know, on my end, I'd been flying back [9:04] and forth between Baghdad and similar [9:06] places in the Middle East. And you know, [9:08] Ben was commuting God knows how many [9:10] flights into Africa. And we looked at [9:12] ourselves and said, "This is the place [9:14] where people live. This is the place [9:16] where people thrive. This is where the [9:19] majority of our lives are spent. And [9:21] then we started thinking about how to [9:22] deconstruct working with American [9:24] cities. And what we came to is at the [9:28] bottom of the pyramid really when it [9:30] comes to how to make cities awesome is [9:32] the idea of safety. The idea that people [9:37] need objective safety and also need to [9:39] feel safe. [9:41] And [9:43] when that stability is there, it's [9:45] amazing what's possible. And there were [9:48] all these kind of contrasting but but [9:50] but actually similar concepts that we [9:53] were wrestling with at the time. We were [9:55] thinking about how to deploy inside of [9:57] institutions and also find ways to [10:00] eliminate the mishandling of information [10:02] inside of really complex organizations [10:05] that have access to super super [10:07] sensitive data. And then if we could [10:09] create an the backbone infrastructure [10:13] level of the modern city, then we could [10:17] help to preserve individuality. [10:19] We talk about privacy and other aspects [10:21] of what it means to be like an [10:22] individual person, [10:24] >> but also bring people together. [10:27] And the idea that we could do both of [10:29] those things, preserve and protect [10:32] individual privacy while also bringing [10:34] people together and help them realize [10:36] that if we're staring at the same same [10:37] information, we kind of want the same [10:39] thing. [10:40] >> Yeah. [10:40] >> Like that's what I find really beautiful [10:42] about cities and I think we're very [10:44] aligned on that. [10:45] >> Totally [10:45] >> beautifully said. How many people said [10:48] no before San Pablo said yes? [10:51] You know, it's such a blur at this [10:53] point, but it's definitely over two [10:56] dozen in my opinion. We started the [10:58] company on February 26th of 2018. We had [11:01] incorporated the company a month or so [11:03] earlier, [11:05] >> but February 26 is the day that the San [11:08] Pablo Northern California Police [11:10] Department [11:12] permitted us to step into the building [11:14] with our badges. We had desk space and [11:17] we had access to information so that we [11:20] could actually get to work. And the [11:21] reason we got there is we started to [11:25] research subject matter experts in the [11:27] art of public safety at the at the city [11:29] scale at the municipal level. And we [11:31] found this amazing article about a young [11:35] upand cominging commander in a local [11:38] police department named Brian Bouar who [11:41] was catapulting up the ranks inside of a [11:43] you know a paramilitary organization [11:45] that's very structured that takes years [11:46] and years to reach higher levels of [11:49] leadership and he was young and [11:52] intelligent and creative. And there's [11:55] this article we found online about [11:57] Operation Red Reach, which was one of [11:59] the most iconic um crossjurisdictional [12:03] gang affiliated narcotics [12:06] investigative, you know, work uh that [12:08] happened in Northern California at the [12:10] time. Um and Brian [12:13] was a covert [12:15] um operator [12:18] trying to figure out how to move this [12:22] stuck investigation forward. M [12:24] >> and I thought that took so much courage [12:27] for someone like Brian to come in and be [12:31] the reason Operation Red Reach and it [12:34] was obviously wasn't just Brian but so [12:36] many people um surrounded that work and [12:39] did something so impactful um and made [12:43] headlines but more importantly um you [12:45] know [12:47] changed the the the state of safety in [12:49] Contraosta County. We just called him [12:52] and said, "Brian, [12:54] we don't know that much. [12:57] >> We may have some utility, [13:00] >> but I don't know. [laughter] [13:02] Can we come in, ask you some questions, [13:04] learn from you, [13:06] try to understand some of the awesome [13:08] things you've done, including some of [13:10] these amazing prior investigations." [13:12] >> Yeah. [13:12] >> And then maybe build something over [13:14] time, and just see what happens, and [13:15] lead with that level of trust. And that [13:17] was what finally got us access. take a [13:20] chance on you. [13:21] >> Yeah, truly. Yeah. [13:25] >> Do you do ride alongs? [13:28] >> We have amazing photos in those days. [13:31] >> Our first ride along was actually with [13:32] Oakland PD which was which was way when [13:35] we were still in this like deep research [13:37] race and that was a really eye opening [13:39] and interesting experience [13:41] >> overnight I think. [13:41] >> Yeah, it was overnight. [13:44] >> Yeah. Yeah. [13:46] >> And it was also uh eye opening. It's so [13:48] I mean the most people that are coming [13:51] um I think from from Silicon Valley [13:53] going on ride alongs are not necessarily [13:56] engineers at the time you know reporters [13:58] or uh other other types of you know [14:00] community interest organizations. And so [14:02] you know I think we we had the white [14:04] glove service at the time [14:07] >> um and it it really was it was eye [14:09] opening that it's so difficult to [14:10] penetrate uh and and build trust with [14:13] these organizations. And so there were [14:14] so many lessons at the time [14:16] >> that signal to us that we just have to [14:17] take our time. [14:18] >> Yeah. [14:18] >> Yeah. [14:19] >> You started this company [14:21] >> 2017 2018. [14:24] >> I think defund the police was, you know, [14:27] was starting to get going maybe in the [14:29] early days of the company. What did that [14:31] feel like? And how did it make you dig [14:34] deep and think about your values? I'll [14:36] share that you know on a personal level [14:38] I think I'm a very um I tend to be a [14:39] very experience-based learner like very [14:42] tactile very you know under want to hear [14:45] the concepts but [14:48] build a bottoms up understanding through [14:50] action [14:51] >> and [14:52] I have found in my life with that [14:54] mentality it fundamentally leads me to [14:56] question assumptions [14:59] >> and um the the intent is not to question [15:03] assumptions in a in a low trust way in [15:05] skeptical way at all times. It's it's [15:07] not, you know, I think there's a danger [15:09] of becoming overly cynical, but um to [15:12] really understand truth and [15:16] love like curious adventure um that I I [15:21] think that is the the kind of that the [15:24] psychological backing that led us to [15:27] build our business in the midst of a [15:29] really tumultuous time in America. [15:31] >> Yeah. you know, the I I have a personal [15:33] affinity for working with people who [15:38] sometimes feel misunderstood, frankly. [15:42] And [15:44] um it just it just is it's it's an very [15:47] intuitive passion for me um to help help [15:52] help them feel more understood. Uh I [15:54] think many people at our company [15:56] probably have have character qualities [15:57] like that, frankly. Um, and when I think [16:02] about moments like [16:06] COVID and the protests that were [16:10] happening at the same time, the [16:11] convergence of, you know, social unease [16:14] in America, [sighs and gasps] [16:16] we were holding a lot of complexity. We [16:18] had very close friends that we've [16:21] graduated with and built things with in [16:23] the past that were um standing outside [16:25] of the women's center in San Francisco [16:27] protesting the police. And [16:31] we were driving a Honda Accord across [16:33] the Bay Bridge out of San Francisco [16:34] every day to work with gang homicide [16:36] investigators [16:38] and we were holding those two truths at [16:40] the same time. And I [16:45] to me [16:48] it it's not it's not enough to to look [16:53] at a very difficult situation from the [16:56] outside. Like the privilege is jumping [16:58] into the pool and swimming. [16:59] >> Um and I think through that process we [17:02] establish empathy. We realize that oh [17:04] this you know this this thing that we [17:07] thought was wrong is actually really [17:08] complicated. there are a lot of gray [17:10] zones and actually the deeper we go the [17:12] more nuanced and textured the problem [17:15] actually is like how empowering is that [17:17] right um [17:19] >> and uh and so in some ways even though [17:22] in that period of time our business is [17:24] much more diverse now right I mean we [17:25] work the the through line is is is the [17:28] is delivering safety and and prosperity [17:31] to cities but at the time we were only [17:33] working with police departments now [17:35] police fire emergency management health [17:36] services you know all all kinds of stuff [17:38] going on but engineers [17:42] would wouldn't [17:44] even respond when we made a recruiting [17:47] [laughter] call. I mean that was it was [17:49] a fascinating time. Uh and I think that [17:51] that level of polarization to me for [17:54] both of us in our lives I think we've [17:55] seen the pendulums shift. Um, and uh, [18:00] we're trying to deliver solutions that [18:03] are apolitical, that are what people, [18:06] communities, groups of people, these [18:08] institutions that serve them truly want [18:10] and need. And in many ways, I I actually [18:13] don't think that's controversial at all, [18:15] uh, if done well. [18:16] >> How is your business structurally [18:18] different from the data collection [18:20] companies like a Flock or an Axon? [18:25] So we've we've been operating in public [18:27] safety since since 2018 and I think most [18:31] of the companies that have come before [18:32] us um have built their business like [18:35] fundamentally on the back of data [18:36] collection. So the idea that um a [18:39] hardware or a software solution [18:42] is installed inside of a customer base [18:44] and is fundamentally about the input [18:47] whether it's passive collection through [18:48] a sensor or you know human beings like [18:50] inputting information directly um the [18:52] collection and storage of that [18:53] information. Historically, those [18:55] companies have grown because once they [18:57] have information that it belongs to the [18:59] customer inside of their system plus [19:01] distribution advantage, they can kind of [19:03] sell more stuff and those the additional [19:06] stuff tends to be more collection [19:07] systems. [19:08] >> Paragrin is like the inversion of that [19:10] entire model. The idea for us is we're [19:13] we're fundamentally in the business of [19:14] joining disparit information to provide [19:18] a more secure properly governed solution [19:21] that sits on top of the pre-existing [19:24] systems that helps people do better work [19:27] helps people get more precision and [19:29] accuracy in the answers to their [19:31] questions. And I think for companies [19:34] like ours being not in the business of [19:39] bringing more data to a customer but in [19:41] the business of building solutions that [19:44] drive greater levels of precision and [19:46] how a human being interacts with their [19:47] data um is a uh kind of like flipping [19:51] the historical model on its head so to [19:52] speak. [19:53] >> Okay. So structurally almost the the [19:55] business incentive is the opposite [19:56] almost. It's not about maximizing the [19:59] amount of data that you're collecting [20:00] and creating network effects of that [20:02] data. It's [20:03] >> I think the world's very concerned about [20:06] the idea of like the amalgamation of [20:08] data and and the kind of like central [20:10] you know whether public sector or [20:12] private sector kind of authoritarian [20:14] body that has access to this privileged [20:17] information and what will they do with [20:18] it and for us we almost want to [20:20] decentralize that to get to a world [20:24] where like we are actually not in any [20:27] shape or form in the business of [20:29] bringing more data to the customer that [20:31] they don't already have. The fundamental [20:33] problem is that they can't utilize the [20:34] data that they have or utilize it in a [20:37] way that's secure and high trust for the [20:39] communities that they serve. [20:41] >> I'll just say from day one we've been [20:44] really focused and thinking about this [20:45] idea of permission controls, data [20:48] governance, sovereignty. These are [20:51] necessary ingredients to kind of deploy [20:54] to these types of high stakes [20:56] institutions. And [20:58] Nick is right when he talks about trust [21:01] and we think a lot about trust. Trust [21:03] both with the community. These are the [21:06] uh these public servants are ultimately [21:08] serving the community members. We have [21:09] trust with with those public servants [21:11] and the community that we have to [21:12] uphold. And so we think a lot about this [21:14] and how we how we build what we build uh [21:18] and and how it interacts with the with [21:20] the end user. [21:21] >> Yeah. It seems like the data governance [21:23] and data being so locked down is almost [21:26] the way to reconcile the tension between [21:29] public safety and you know not becoming [21:32] the surveillance state like that that is [21:34] the answer kind of this [21:36] >> idea of you have local intelligence [21:38] local data everything's locked down. Um [21:41] actually maybe share a bit about your [21:43] philosophy on data data ownership um and [21:47] data sharing. Yeah, we we deeply believe [21:51] in the idea that each customer, each [21:55] institution owns their own data. This is [21:58] ultimately the [22:01] that organization is serving the [22:04] community. That is the community's data. [22:06] That is the organization's data. It's [22:08] not Paragan's data. And we think about [22:11] providing the the controls and uh [22:15] capabilities to allow them to [22:19] very securely roll this out within their [22:22] own department but also share this with [22:25] uh select pieces of information when [22:27] they need to. Without these types of [22:30] solutions, you get folks, you will put a [22:32] bunch of data in the back of a car and [22:34] drive it across uh a city and there's [22:37] there's there's actually less control [22:38] there. [22:39] >> Uh and so we think a lot about where [22:41] where can we actually enable [22:44] the outcome for our user and provide the [22:47] controls that are specific and robust [22:50] that enable them to not overshare, share [22:54] too much, and that they can trust in [22:55] what they're what they're actually [22:57] delivering. What are the biggest use [22:58] cases you see uh for your customer base? [23:01] Maybe what are people starting to play [23:03] with now and then where do you see that [23:05] going? [23:05] >> Yeah, this is what excites me the most. [23:07] I'm [23:08] >> I [23:09] >> Me too. [23:09] >> Yeah, [laughter] I think I think I think [23:11] when you initially do deploy this type [23:12] of AI to these organizations, you often [23:15] just get what I'd say a nice search. you [23:19] know, maybe you were looking for address [23:21] and now you don't have to look at a [23:22] bunch of rows and you see everything [23:24] that happened that address in a nicely [23:26] formatted way. I think the as we have [23:30] had users get more used to the product [23:32] and understand it better and how it [23:34] works, you start to see the floor get [23:36] raised for everybody across the [23:37] department, you start seeing these [23:38] really interesting, deep types of [23:40] analysis that previously were just [23:43] impossible. So, I'll give a couple [23:46] examples here because uh this stuff's [23:48] really interesting to me. We're working [23:50] with this county in Florida and uh the [23:54] other month they had to do over 100 [23:56] water rescues and they're like, "Why? [23:59] We've never had to do this many water [24:01] rescues before." [24:02] >> Water rescue is is a when there's a [24:04] flood, rescuers need to get on a boat [24:07] and literally go out and rescue someone [24:09] from their car that might be stalled in [24:11] the middle of the flood or their home. [24:12] >> Yes. [24:13] And so they started they started asking [24:16] paragrin and started interrogating this [24:18] question and uh what what they what they [24:22] started to uncover is that okay these [24:24] weather patterns they've happened before [24:26] uh but after the a couple iterations and [24:29] the agent doing some deep research it [24:32] came out with the the pattern that oh [24:35] these weather patterns have never [24:36] occurred for three consecutive days and [24:40] those types of weather patterns they [24:42] create these sand channels. And those [24:44] sand channels are the perfect conditions [24:47] for rip currents. And those rip currents [24:50] cause a lot of issues for people who are [24:52] in the water at that time. And uh that [24:56] is actionable for the agency and they [24:58] can think deeply about that [25:00] >> and and overlay that you know their [25:01] terrain maps and all of the unique [25:03] things that are that are count. This is [25:05] all fed by a bunch of like they [25:07] obviously have to have incidents [25:09] information, 911 calls, they have to [25:11] have weather information. All that [25:12] information is integrated and and kind [25:14] of at uh at the disposal of the AI to [25:18] help the the the customer. [25:20] >> Yeah. [25:21] >> That and we just can't I mean there's no [25:23] way we could have made this up, right? [25:25] It's like we if we were sitting in our [25:27] headquarters [25:28] >> trying to pontificate [laughter] [25:30] >> on what you know emergency responders in [25:32] a in a hurricane needed. Like there's [25:34] zero chance we would have figured that [25:35] out. [25:36] >> Super interesting. Another I mean [25:37] another one that I'm like excited about, [25:39] you know, there was a uh detective who [25:42] was investigating a threat against a [25:45] synagogue and the uh what he was trying [25:49] to do was trying to figure out were [25:50] there any other anti-Semitic threats [25:52] that had happened to these two [25:54] synagogues in the area. Now, when you [25:56] think about that question, uh, and [25:58] you're a detective, what are the key [26:00] words that you search for to try and [26:02] find that information? It's actually [26:04] very difficult for for keywords to to [26:06] find that. And what gets me uh excited, [26:09] this is another area that's just [26:11] previously impossible. And with the help [26:14] of AI, they're able to semantically [26:16] understand a lot of the information, all [26:19] that they have access to. and they're [26:21] able to actually pull out and they were [26:23] able to find a bunch of pattern of these [26:25] threats that were occurring against [26:27] these these synagogues and I thought, [26:29] you know, that that was that was pretty [26:31] interesting to me and a novel use case. [26:33] Yeah. [26:34] >> Interesting. Okay. So, you got semantic [26:35] search and embeddings going on. You got [26:38] um sounds like reasoning models that [26:40] uncovered the use case. Um, so one of [26:43] the most interesting thing is is because [26:45] you've brought in all this context and [26:47] you've stitched it together, you can [26:49] kind of let the AI do its thing. [26:51] >> Certainly 95% of the work is what [26:55] happens before the user types in the [26:57] question. What is all the preparation [27:00] that you do to get to a place where the [27:03] AI can [27:04] >> answer accurately, site those questions [27:07] accurately? That's a really hard [27:08] problem. We spend a lot of points on [27:10] that. We have a lot of agentic use cases [27:13] and I you know our a lot of our [27:15] engineering points are spent on that [27:17] data preparation getting that that AI [27:18] ready. There's also this idea that AI [27:20] allows you to write software rapidly and [27:23] what happens when the cost of generating [27:26] the software is virtually zero. And what [27:28] if we could apply uh uh provide the [27:31] platform for our deployment team that [27:33] gives [clears throat] them the security [27:35] and governance controls that gives them [27:36] the APIs they need and then they're [27:39] allowed to [27:40] >> write the world with software. And so [27:43] what we're seeing is that these [27:44] deployment strategists, you know, not [27:45] every customer needs a a an agent. Maybe [27:50] they need a hurricane simulator. And the [27:52] way they're able to write this really [27:54] well and is with is with AI and and I [27:57] think this rapid innovation cycle is [27:59] just super interesting to me. So [28:00] >> so cool. Is it your deployment [28:02] strategists that are kind of coding [28:05] things in customer environments or do [28:07] you think your customers might even you [28:10] know they they have all these ideas in [28:11] their head of what they they want. Can [28:13] they go from idea [28:15] >> to working application themselves in [28:17] your environment? [28:18] >> 100% we enable that. I think the you [28:22] know what we see most often is a [28:24] partnership with our deployment team to [28:28] to provide that kind of technical [28:31] expertise and be able to take the idea [28:34] and bring it to fruition. We do have [28:36] some customers that are able to to kind [28:37] of leverage the platform in that way. [28:39] But a lot of that motion comes with our [28:41] with our deployment team. [28:43] >> Do you think it'll change over the [28:44] coming years? [28:45] >> I see it changing right now. I mean I'd [28:47] love to get your thoughts on on the [28:48] spend. I I think it takes confidence to [28:52] embrace that and say you know what like [28:55] let's let's innovate and you know may [28:58] may the best method win. So much of what [29:01] paragrin does and it on the technology [29:04] side is emergent. Paragan has been [29:05] providing domain expertise and for our [29:09] customers we deeply understand and [29:11] empathize with law enforcement, fire [29:14] departments, EMTs [29:16] and then we provide technology to enable [29:19] them to achieve their most important [29:21] missions. And that technology can change [29:24] over time and that's like the great [29:26] thing about it. Most of that stuff is [29:28] downstream of really high quality data [29:32] and you can build charts and that chart [29:37] could be completely useless if it's not [29:39] accessing the right information and [29:42] accurate information. The same is true [29:44] with agents and the same will be true [29:46] with the next technology. And so I I [29:49] believe like the core strategy of [29:51] Paragon has remained very unchanged in [29:53] the in this new age of of AI and it's [29:55] just we are applying a new technology to [29:58] see how it impacts our customers. [30:00] >> Are you doing anything on the long [30:01] horizon agent side or background agents? [30:04] >> Yeah, absolutely. I a couple couple [30:06] things here. So I think stepping back a [30:07] little bit like what is the paragrin [30:08] platform? [30:11] We think about 50% of our engineers [30:14] spend most their time working on the [30:16] data platform that enables our [30:19] deployment team to integrate data uh [30:22] that the that the customer already owns. [30:25] >> And we've at this point integrated, you [30:28] know, tens of thousands of data sets [30:30] across all these customers and we have [30:33] built uh an agent that enables our [30:36] deployment strategy team to integrate [30:38] this data agentically. And I think at [30:40] this point we have this amazing eval set [30:43] where we can you know deterministically [30:46] uh evaluate these agents for [30:48] completeness and correctness on on [30:49] integrations. At this point we're seeing [30:51] about 90% of our we use a version of of [30:55] Python notebooks to do a lot of our [30:58] integrations. About 90% of that is [31:00] written by agents with the oversight of [31:03] our um of our deployment team. And these [31:06] agents run for hours, right? and they [31:08] will they will you know analyze look at [31:11] the databases understand the ontology [31:15] start to piece together different pieces [31:17] that need to be integrated those split [31:19] off into sub aents that all do a bunch [31:21] of work communicating back to kind of [31:22] the orchestrator agent [31:24] >> that's one bit on kind of uh long [31:28] horizon agents [31:29] >> reminds me of how people are using these [31:30] coding agents for like codebased [31:32] migrations also similarly just [31:34] longunning unglamorous work [31:36] >> 100% I love problem and because it's [31:38] it's verifiable and that makes the [31:40] problem a lot easier. This is why coding [31:41] agents are in a lot of ways a lot [31:43] easier. And so the second piece that we [31:46] spend a lot of time on is [31:48] >> what do our agents look like for [31:50] operational outcomes for our end users. [31:53] And the very first agent that we built [31:56] was a cold case agent. And uh this [32:01] agent, you know, just give you some [32:03] context on what some of these very [32:06] high-profile investigations look and [32:09] feel like. You are uploading 200 to 300 [32:13] gigabytes of data. These are videos, [32:16] they are audio, it is images, it is a [32:20] ton of uh PDF files. and [32:25] the detective is tasked to with going [32:28] through all that. So that that that in [32:29] and of itself takes a really really long [32:31] time. I've been in these police [32:32] departments where they actually have [32:33] paper records of all these and they are [32:35] boxes that are like this big and you [32:38] know they have all these CDs attached to [32:39] it. It it's really an outrageous amount [32:41] of data. Mhm. [32:42] >> And so our cold case agent uh we first [32:45] did we built this with customer and [32:47] again like this comes back to like how [32:48] we like to how we like to build and we [32:50] were working with a a customer who had [32:52] worked a case where uh a man was uh [32:55] wrongly convicted and uh they were able [32:59] to exonerate this individual [33:02] >> um and they said, "Hey, can you [33:04] reproduce this result with an agent?" [33:07] And so we took all of the evidence and [33:10] data that they had on that case and [33:12] started to work on an agent that would [33:16] run for 30 minutes, 60 minutes to start [33:19] to glean insights and eventually we got [33:22] to the place where it could reproduce [33:24] the results that those detectives had [33:26] gotten to. So that was that was where [33:28] that was like kind of the the [33:29] manifestation of our first kind of cold [33:31] case agent. Now we're using this in a in [33:33] a few departments across uh the US. We [33:36] recently were working in a county in [33:38] Wisconsin where [33:40] uh this again similar type of case 300 [33:43] gigabytes of data and they were able to [33:45] identify and place a a the suspect at [33:50] not just the scene of the crime but [33:52] where the body was found. And this was a [33:56] single there's about a few cell records [33:59] cell call detail records kind of ping of [34:01] a phone that was scattered amongst Yeah. [34:05] a lot of data. And so this this I this [34:08] was really interesting to me and I think [34:10] >> points to this area where we could [34:12] really uplevel and help our our you know [34:18] public servants [34:19] >> unbburden them with all this [34:20] administrative watching hours and hours [34:22] and hours of video and listening to [34:24] hours and hours of audio. [34:25] >> As a quick comment this this example Ben [34:27] bro is so impactful to the organization [34:31] as a culture for our business. we have [34:33] the way that we maintain trust is not [34:35] actually taking credit and shouting from [34:37] the rooftops about the awesomeness of of [34:39] of what happened here. Um I think that [34:42] is one of the fastest ways to break [34:44] trust frankly with these organizations. [34:46] The idea that you know we're going to [34:47] scoop up the work and trumpet our skill [34:51] and the way that we impacted the world. [34:52] Uh so I think being the quiet [34:54] professionals in a context like this um [34:57] and empowering the the customer is why [34:59] we have access to the next problem. I [35:02] wish you would talk about it more though [35:03] because I think that we're in this [35:06] moment where public distrust in AI is so [35:10] high. [35:11] >> Yeah. [35:11] >> And this is a wonderful story, right? I [35:15] think that coming back to an earlier [35:17] example of um uh you know trying to [35:19] identify threats to a synagogue um uh I [35:23] think about network effects and I think [35:26] about um what motivates us to tell our [35:30] story. Um I um I think the convergence [35:35] of this idea of a kind of a a [35:37] centralized authorit authoritarian like [35:39] AI capability and surveillance uh in [35:43] American uh or or you know national [35:46] society. The convergence of these forces [35:48] has created this immense pressure and [35:50] and I think distrust um in in many of [35:54] the the the organizations and and [35:56] frankly the communities that that we [35:58] work with. Um [36:00] >> I that can't be ignored first of all. Um [36:03] and I I think if we look at the the [36:06] organizations we work with and also the [36:07] data landscape, the network effects are [36:10] there and they are unbelievably strong. [36:13] Like for example, it would be very easy [36:15] for an organization that's trying to [36:16] kind of hack their way to building [36:19] something useful to go and hit open [36:22] source and pull in data that they might [36:24] not. there might be a gray zone on [36:26] whether they can or should access that [36:29] based on laws, regulations, ordinances, [36:31] standard operating procedures of of the [36:33] department. And for us, like we have to [36:37] protect against those moments where the [36:40] it's not like we don't need to we don't [36:43] need to break the rules, nor obviously [36:45] should any organization break the rules [36:48] in order to find a way to like introduce [36:50] AI or technology into these complex [36:52] environments. The opposite is true. We [36:55] have to protect the data, right? And we [36:57] have to protect these organizations that [36:59] by virtue of their structure have [37:02] astounding network effects, right? It's [37:04] almost like the anti-etwork effect [37:06] proposition. Like you how do you [37:08] preserve the sanctity of the data [37:12] and the ownership of the data and the [37:14] ways of working for every individual [37:16] agency, every individual organization [37:18] and then building the connective tissue, [37:20] building the interoperability [37:22] is going to happen. [37:26] >> But and this is where things that might [37:27] sometimes seem mundane but we find [37:29] really fascinating like how do you think [37:31] through data governance? How do you [37:32] think through permissioning logic in the [37:34] context of AI become unbelievably [37:36] important? Um so I'm feel really [37:40] motivated uh to talk about that [37:43] >> in particular. [37:44] >> Paragrin the anti-et network effects [37:46] business. [37:48] Um, I think it's important, especially [37:50] given the heat now around [37:54] AI is so powerful and so a lot of this [37:57] data that used to exist, I think people [37:59] now fear what what happens when it's all [38:03] >> swooped up into a central panopticon. [38:06] That's deeply unamerican. [38:08] >> Yeah. Yeah. Yeah. [38:09] >> Um, [38:09] >> yeah. Yeah. And I think that as a [38:11] business, it's like I find that our [38:13] business is a big practice in letting [38:16] go. It's how do we build high integrity, [38:19] transparent solutions for our customers? [38:22] How do we make that transparent [38:25] >> to our customers and their constituents? [38:28] >> Um and then how do we [38:33] effectively [38:35] let let go and not try to um grow too [38:39] fast? [38:40] >> Are there technology decisions that you [38:41] have to make that are morally nuanced? [38:44] So, for example, facial recognition. I'm [38:46] curious your stance on that. And then [38:49] more generally, what what is your [38:50] northstar for hard decisions? [38:52] >> Yeah, quickly on facial recognition. Um [38:55] I I think the idea of like a Silicon [38:57] Valley company imposing um a decision uh [39:01] that is kind of general purpose for an [39:04] industry is fundamentally wrong. I think [39:06] the idea that we as an institution would [39:09] assert whether it be you know the [39:11] utilization of a technology or a [39:13] retention policy I mean it's um it's [39:16] it's very much not how we we think. Um [39:19] for instead uh the idea is to to help a [39:25] customer understand the context in which [39:28] they're operating and actually bring to [39:30] light all of the considerations they may [39:32] or may not know and then help them and [39:35] this requires patience. Help usher in [39:38] the right way of doing business, the [39:40] right way of thinking about deploying [39:41] these technologies on facial wreck. Most [39:44] public safety agencies in America ch and [39:46] their communities choose not to [39:49] implement facial recognition. Um some [39:51] take a very strong stand and and some [39:53] are are uh uh it's it's just a you know [39:56] a matter of like subjective preference [39:58] that they'll assert. Um but uh um the [40:01] idea of creating you know a [40:03] technological um red line uh without [40:06] understanding the texture and context is [40:08] is not not a boundary line that we think [40:11] that we can assert on top of our [40:13] customer. [40:13] >> So follow the customer, follow the law. [40:15] >> Yeah. [40:16] >> And and expose it and and and create [40:18] clarity around, you know, what leftright [40:22] limits and and what what the [40:24] institutional norms may be. What I mean [40:26] it's astounding how cities call each [40:27] other, [40:28] >> right? they call each other for advice. [40:30] And so if you really just listen to the [40:32] way that people learn inside of the [40:34] industries that we serve, what we [40:36] realize is um helping them streamline [40:39] the way that they can get the best [40:40] possible information so they can make [40:42] their best decision even about things [40:44] like what technologies to use or not [40:46] goes a really long way. Maybe in a [40:49] similar vein, I want to talk about [40:51] delivering technology to the underdogs [40:53] and how you're able to scale this very [40:56] deep customer motion into a very [41:00] different customer base than, you know, [41:02] the other company that notoriously has [41:04] scaled this motion, Palunteer, which [41:06] notoriously doesn't take anything less [41:07] than eight figure contracts. [41:10] >> Um, [41:11] >> did your mentors try to convince you not [41:13] to do this? Did people tell you like the [41:14] economics of this are not going to [41:15] scale? And what made you think it was [41:17] going to be possible to deliver [41:19] technology to the underdogs and serve [41:20] them in this way? [41:21] >> I think it takes confidence. I think it [41:23] takes confidence to to believe that if [41:26] you build something well, you can solve [41:29] problems that have never been solved [41:30] before that will scale later. M [41:33] >> and the idea of going deep to build a a [41:38] vertically integrated tech stack that [41:41] everything from network access to the [41:43] permissioning and the governance and the [41:44] ETL and the pipeline and the ontology [41:47] and the UX and the APIs. It's a to [41:50] configure all of that and create a [41:52] system that's open interoperable and can [41:54] check these boxes. It it the audacity I [41:58] think is is is real. Um and and to build [42:02] each of those components in a way that's [42:04] first class, not just a check the box. I [42:07] mean, it's we were [42:09] >> truly thinking about how to deliver [42:11] these types of technology solutions to [42:13] an end market that has never been able [42:15] to use these solutions before, [42:18] >> let alone at the price point that we [42:19] were able to to deliver them. I think [42:21] the the uniqueness about paragrin is for [42:24] each organization we'll take the time to [42:26] deliver solutions all the way to the [42:29] outcome and then we'll own the [42:31] responsibility with our customers of [42:32] getting to that outcome which might seem [42:34] like extremely manual extremely [42:36] laborious extremely unstable or or [42:39] unscalable [42:40] >> and [42:42] >> what we found is being able to do that [42:44] >> resulted in a technology platform that [42:48] could show up and do something very [42:50] different for these organizations [42:53] and also scale. [42:54] >> I've had the chance to talk to a few [42:55] deployment strategists at Paragrin and [42:57] they are they're amazing. They're so [43:00] smart, so humble. They so deeply embody [43:03] the mission and so [43:04] >> I think you guys have selected for and [43:06] trained a really extraordinary group of [43:09] people. [43:10] >> How do you do it? What do you look for? [43:12] What do you what do you what do you [43:13] interview for? How do you train these [43:15] people up? First of all, I mean, I it's [43:17] it's the great it's the greatest [43:18] compliment because I and we love our [43:20] team. When you can build a culture [43:23] where people are trying really hard and [43:27] can achieve excellence, it's almost like [43:29] how could you not create the [43:30] environmental factors and the the [43:34] feedback culture, which sometimes isn't [43:36] easy, right? Sometimes it's it can be [43:38] really intense. Ben and I, I would say, [43:40] you know, we've aligned we've taken [43:41] tests and we've aligned on some [43:42] character qualities. One of which is [43:44] definitely the the pursuit of [43:45] excellence. I think a lot about these [43:47] environmental factors of um uh like [43:50] trusting the individual contributor. [43:53] It's a very um important thing that [43:56] we're trying to hold on to. And to do [43:58] that with a forward deployed engineer [43:59] actually to truly do it really requires [44:02] trust because you we use this analogy of [44:04] a a dark cave, right? will will send a [44:08] person into a cave and they have a tool [44:11] belt and hopefully the tools are pretty [44:13] good, right? And and there might be a [44:14] rope, you know, with a friend at the the [44:16] entrance of the cave that, you know, if [44:17] they yell for help, you can you can kind [44:18] of pull them out. But truly, we are [44:20] sending people into these zones [44:24] and we're saying maintain your integrity [44:28] like maintain your principles. [44:31] Here are the tools. Go do good. And to [44:34] do that at scale, I really believe [44:36] requires um [44:39] institutional trust and and a belief [44:41] system that innovation actually happens [44:44] at the farthest fringes of our [44:45] organization. That the way that our [44:47] engineering product design organizations [44:49] are fed [44:51] >> absolutely [44:51] >> is through that innovation process [44:54] across the country and and now now [44:56] across multiple countries with our with [44:58] our forward deploy team. [44:59] >> Yeah, [44:59] >> it's fascinating. so different from when [45:02] I talk to companies who view forward [45:04] deployed as a cost center. Uh you guys [45:07] are [45:08] >> it's fundamentally very different [45:09] talking to you. [45:10] >> Definitely R&D. Definitely R&D and and [45:13] growth. I mean that we there after we [45:15] land inside of a customer base, we have [45:17] these models where we'll we'll do these [45:19] within customer pilots for all intents [45:21] and purposes where we'll keep sprinting [45:23] on additional use cases. And in many [45:26] ways the you know how we land and then [45:28] how we expand in in a customer base is [45:30] fundamentally about um leading from the [45:33] front through our forward to plate [45:34] engineers. [45:34] >> Have you discovered any new products [45:36] through this motion? [clears throat] [45:39] >> 100 I mean 100%. It's it's just the [45:41] whole thing like I think it's [laughter] [45:43] hard to even separate that. I mean, I [45:45] remember walking or walking uh driving [45:48] uh across the Bay Bridge, San Pablo, [45:51] typing on the computer and Nick's like, [45:52] "Okay, we got to get these reports [45:53] done." And like this is and this is what [45:55] this detective needs and literally [45:58] typing as fast as I could to try and [46:00] code what we needed at the time and when [46:03] you're And it wasn't theoretical. It was [46:05] Aaron Blazale who had an extraordinarily [46:08] urgent request. [46:09] >> Yeah. And and [laughter] like I feel it. [46:10] And so it's so hard to separate. Very [46:12] rarely are we in a room pontificating [46:15] about you know what to build. [46:17] >> I I really think about the FD motion and [46:20] how it impacts products our product in [46:23] in two ways. [46:25] >> One is [46:27] kind of what are the primitives basic [46:29] primitives that the technology platform [46:31] needs and there's a couple of examples [46:32] of this that I think are are fascinating [46:34] and reflect the ingu ingenuity of our [46:37] forward deploy team. For the longest [46:39] time in Paragan, we had no ability to [46:42] edit specific fields and properties. And [46:45] so one forward deployed engineer to get [46:47] around that made an integration based on [46:51] the comments people would write on these [46:53] objects and that comment would be read [46:56] by Paragrin and then update a property [46:58] and that was the way they implemented [47:00] editing. And so we saw that we're like [47:02] we need to implement [laughter] editing. [47:04] And by the way, the reason that teammate [47:06] did that is because our intent was and [47:10] and this is the way that the coaching [47:11] and leading happens. It's like your job [47:14] is to hit the objective like to hell [47:16] with the technology. [47:17] >> Yeah. [47:18] >> I mean, our our our job is to build the [47:20] technologies and the tools that empower [47:22] you to do higher and higher levels of [47:24] work, but your job is to achieve. [47:26] >> Yeah. [laughter] [47:27] And that's what creates these crazy [47:28] things that might look very unscalable, [47:30] very hacky, but ultimately are the [47:32] absolute best signaling network. [47:34] >> It's great signal for what actually [47:36] works. And it's it's like the [47:38] quintessential rapid prototyping. [47:40] >> So So that's like category one. You get [47:42] these these product primitives that you [47:44] know that you need to provide the team [47:46] in order to to achieve the objective [47:48] they're trying to achieve. [47:49] >> [gasps] [47:50] >> I think like category 2 is almost uh [47:54] what I call like innovation lab where [47:58] the deployment strategist goes and [48:00] builds something totally unique that [48:02] we're probably not going to integrate [48:03] back into the product just because it's [48:06] it's such a unique capability [48:08] specifically for that customer. And this [48:10] stuff gets me really excited and I I [48:12] love watching and seeing what our [48:14] deployment team builds. was just looking [48:16] at one the other day where someone had [48:19] built a hurricane simulator [48:21] >> in Parag and I was like, "How did you do [48:23] that? [laughter] I didn't know you could [48:25] do that." And so that was that was [48:27] really interesting. There was another [48:28] deployment strategist who built this [48:31] thing where you could you could it it it [48:34] integrated a bunch of data from all [48:35] these different sources, 911 call times, [48:37] like budgets, and you could place a fire [48:40] department in in the city and it would [48:43] give you an estimate on uh how many [48:46] people that would impact. [48:47] >> Unbelievable. And she was at the company [48:48] for a month. [48:50] >> Maybe last last set of questions. Uh 10 [48:52] years from now, suppose everything's [48:55] gone right. Paragrin is the [48:57] institutional memory layer [49:00] >> for 10,000 cities. [49:02] >> Um, I think that's great power that [49:06] outlasts any one administration and with [49:08] great power comes great responsibility. [49:11] How do you think about that? [49:13] >> I think there's the outcome and then I [49:15] think there's the path to get there. M [49:17] >> um the end game of 10,000 cities [49:22] um [49:23] uh I think requires an operating model [49:26] that is fundamentally um about uh [49:29] infrastructure actually it's like we as [49:31] a technology organization are delivering [49:33] technology infrastructure that empower [49:36] these organizations to do with their [49:38] data as they [49:41] would like and as they are required to [49:43] do. M [49:45] >> and every city, every jurisdiction I [49:47] deeply believe is unique and I think the [49:50] idea of um preserving that uniqueness is [49:55] actually beautiful [49:58] >> and so I'm really excited about that. [50:01] [laughter] Like how do you make how do [50:02] you make each you know the poperri of [50:04] cities make each one more awesome in [50:07] their unique way? [50:08] >> Um that's cool. Uh [50:11] the path to get there requires serious [50:15] integrity, moral compass, core values um [50:19] not just in words but but in um our ways [50:22] of working, our actions and that never [50:26] goes away. It never goes away. the way [50:29] that we think about the next marginal [50:32] customer that we support um requires a [50:36] level of um handholding and delicacy [50:39] that candidly I don't think was possible [50:41] even 5 years ago because the marginal [50:43] cost of doing what we do to drop it [50:46] below a million bucks a year I mean it's [50:48] radical right coming back to the origin [50:50] stories [50:51] >> in many ways you know supporting state [50:54] county and city level public safety was [50:56] you know there were there are a a [50:58] lineage of failed business units of [51:00] major astounding organizations that [51:03] tried to do it and failed because to [51:05] deliver these solutions in a way that's [51:06] tailored at a price point that these [51:08] organizations can afford was never [51:09] possible. And so to drop the to create [51:12] the efficiency to drop the price by an [51:14] order orders of magnitude is um has has [51:17] earned us the right to to try to deliver [51:21] on what you just said. It's like who are [51:22] who are we to to think that we hold any [51:25] power over that? I mean, it's the [51:27] institution that has the power. [51:28] >> Go make every city awesome. [51:30] >> I love that. [laughter] [51:32] >> I think you're in a very very serious [51:33] seat and just in this course of this [51:37] conversation, it's clear that you're [51:39] approaching this with with great care [51:41] and great stewardship. So, thank you for [51:43] what you do. I think the world needs [51:45] more examples of AI being used uh to [51:49] improve lives, to improve communities, [51:50] and um thank you for what you're doing, [51:53] and thank you for joining the pod today. [51:55] Thank you, Sonia. [51:58] [music] [52:05] [music] [52:11] [music] [52:20] >> [music]