Valon Offloads $200B Servicer to License AI Core to Legacy Rivals

Valon剥离2000亿美元服务业务，向传统对手授权AI核心系统 颠覆大型机时代的金融基础设施需要初创企业先行作为受监管的服务商运营，借由三倍的效率提升迫使规避风险的传统企业放弃老旧系统。 Valon创始人Andrew Wang与Linda Du阐述了他们如何系统性地从零重构房贷服务基础设施，放弃单纯的SaaS模式与私募股权并购整合，选择成为一家全新受监管的服务商。通过将盘根错节的州与联邦法规转化为基于事件的记录系统，Valon实现了70%至80%的营业利润率，并在将服务业务出售给Carrington之前，在数十万笔贷款上验证了其架构。全面转型为面向全行业的操作系统后，该公司目前正整合AI智能体、对话式交互界面以及定制评估工具，以实现长尾工作流的现代化。对话还重点提及了Valon与Rithm/Newrez转移400万笔贷款的里程碑式合作，并将大规模软件部署定性为核心系统工程与企业变革管理的精细融合。 Valon通过作为全栈房贷服务商运营来应对州和机构的许可壁垒，在将服务业务出售给Carrington以彻底转向软件授权之前，其未偿还本金余额规模已扩张至约2000亿美元。 (20:00) 作为独立软件供应商推出业务后的六个月内，Valon签署了超过2亿美元的商业合同，其中以Rithm Capital承诺向该平台迁移400万笔贷款最为瞩目。 (33:38) 通过用现代基于事件的账本取代大型机时代的架构，Valon将服务业务从行业盈亏平衡的基线提升至70%至80%的营业利润率。 (20:00) 受监管行业的企业级部署本质上是变革管理问题而非纯技术障碍，需要能够驾驭复杂内部利益相关者诉求的系统思维者。 (37:46)

Valon founders Andrew Wang and Linda Du argue that modernizing the $13 trillion mortgage servicing sector required building an independent, full-stack servicer from scratch to prove operational safety and efficiency before pivoting into enterprise software. Legacy Mainframes and Compounding Data Failures (0:00) Mortgage servicing forms the back-office plumbing for roughly $13 trillion in consumer debt, yet the broader sector remains dependent on a single incumbent whose core architecture was drafted during the mainframe era of the 1960s. Wang observed that legacy accounting ledgers routinely failed to tie out with actual bank cash movements, forcing institutions to rely on ad hoc patches. Because these systems are architected around static state snapshots rather than dynamic event histories, vital historical context is severed whenever a loan transitions, such as during a borrower's death. This creates severe downstream consequences, including heirs being blocked from tax records by call center staff. The Decision to Build a Full-Stack Servicer (10:00) Rather than selling unproven B2B software to resistant incumbents or acquiring legacy servicers burdened by technical debt, Valon opted to launch a de novo regulated servicer. This approach required clearing complex licensing hurdles over three to five years, including securing broker exemptions in California, satisfying New York regulators across a three-year review, and translating overlapping federal statutes—such as RESPA, TILA, FDCPA, GLBA, and TCPA—directly into software logic. Acting as its own servicer allowed the company to survive audits from Fannie Mae, Freddie Mac, Ginnie Mae, and state agencies, proving the software's regulatory compliance before offering it to third parties. Economic Capitulation and Software Transition (20:00) While traditional servicing runs near break-even margins, Valon reached operational margins of 70% to 80%, roughly three times the industry's standard efficiency. To win institutional market share, Valon passed these savings directly to asset managers in the form of lower fees. Having scaled Valon Servicing to approximately $200 billion in unpaid principal balance, the founders executed their planned transition by selling the servicing entity to Carrington and licensing their platform, Valon OS. Within six months of launching pure enterprise software licensing, Valon secured more than $200 million in commercial contracts, including an agreement with Rithm Capital and Newrez to transfer 4 million loans onto the platform. AI Orchestration and Change Management (23:45) Valon utilizes its underlying system of record to deploy agentic workflows that address complex operational situations, such as disaster relief assistance and automatic extended escrow amortizations. Unlike sandboxed coding tools, enterprise mortgage AI must enforce deterministic correctness around legal liabilities, relying on continuous evaluations and fine-tuned models designed to avoid hallucinations. However, Du notes that enterprise deployment has shifted from a pure software challenge into a change management discipline, requiring teams to align disparate corporate incentives and retrain thousands of enterprise staff to achieve real-world adoption. Wang and Du maintain that their operational model extends beyond residential mortgages into commercial real estate and healthcare revenue cycle management. A core ongoing test for Valon is whether its continuous model evaluations and change-management playbooks can maintain strict regulatory compliance while migrating massive multi-million loan enterprise portfolios.

Valon 联合创始人 Andrew Wang 与 Linda Du 阐述了其重构全美 13 万亿美元房贷服务业的战略：团队放弃向传统机构直销软件的常规路径，通过从零申请牌照建立端到端房贷服务商以验证技术与合规性，在将运营利润率提升至 70%–80% 并积累约 2000 亿美元贷款规模后，将自营业务出售给 Carrington，全面转型为向行业授权底层操作系统 Valon OS 的核心软件商。 遗产架构缺陷与房贷服务的人道代价 (5:00) 房贷服务业支撑着全美约 13 万亿美元的消费者债务，但其核心系统长期被单一老牌厂商垄断，底层代码架构多源自 1960 年代的大型机时代。Andrew Wang 指出，传统系统依赖“静态快照”记录当期借款人状态，而非按时间顺序记录完整生命周期。当借款人遭遇失业、重病或身故等变故时，系统往往将历史数据切断并归入孤立的侧库。这导致继承人在接管还款并处理遗产税时，因姓名未列于原始借据而无法调取历史账单；客服人员受限于陈旧终端无法查看流转脉络，迫使家属在悲痛中反复申诉，造成了严重的人道与管理摩擦。 此外，近年来极端天气导致房屋保险费剧增，常造成托管账户（Escrow）大额资金缺口。尽管联邦法规与二级市场投资者准则明确允许服务商将补缴期放宽至 36 个月以平摊月供压力，但 1960 年代架构缺乏自动化规则引擎，只能僵化地默认按 12 个月追缴，人为诱发月供激增乃至违约。旧系统若想延长分摊期，必须依赖人工工单与层层主管审批；而现代底层系统则能自动捕捉保费上涨的诱因，并即时完成合规展期，防止借款人陷入支付休克。 从零构建全栈服务商与法律代码化 (10:00) 面对壁垒森严的规管环境，Valon 权衡了三种路径：向现有服务商销售纯 SaaS 软件、通过私募股权收购旧服务商进行模块化改造，或是从零建立受监管的服务商。Linda Du 解释称，纯 SaaS 路线会被迫迎合传统客户陈旧的作业习惯，极易沦为“十年后仅给旧系统套上漂亮界面”，而收购整合则伴随严重的知识产权纠纷与历史技术债务。因此，团队选择了一条从单笔贷款做起、完全不借鉴旧代码的全新构建路径。 启动独立服务商需要突破层层互锁的监管资质考核。由于房利美（Fannie Mae）等联邦机构通常要求申请者具备盈利历史，Valon 选择从仅需州级审批的非合格住房抵押贷款（Non-QM）切入，并借由加州房产经纪人豁免渠道破局。单是获取纽约州的牌照就耗时三年。与此同时，团队在疫情初期耗费数月将数千页复杂的联邦与州法律逐字解构，涵盖 RESPA、TILA、FDCPA、GLBA 与 TCPA 等重叠法案，将其抽象为可由软件执行的底层领域模型与严密逻辑架构。 验证单位经济与剥离自营服务业务 (20:00) 传统房贷服务业长期处于微利甚至盈亏平衡边缘。Valon 作为自营服务商入局，凭借软件驱动的自动化能力实现了传统模式约 3 倍的运营效率，将营业利润率提升至 70%–80%。Valon 将节省下的成本以低费率形式返还给对经济利益极度敏感的资产管理者，同时以自身系统连续六至七年经受州监管机构以及房利美、房地美（Freddie Mac）和吉利美（Ginnie Mae）的严格实操审计，彻底消除了机构客户对采用未经验证核心软件的安全疑虑。 尽管 Valon 自营的贷款未偿本金（UPV）规模迅速攀升至约 2000 亿美元，足以支撑一家独立的高盈利实体，但管理层严格执行了自 A 轮融资时便确立的“主计划”：将自营服务业务 Valon Mortgage 出售给 Carrington，全面转型为向整个行业授权操作系统 Valon OS 的技术商。Andrew Wang 强调，将技术红利封存在自营服务商内部固然盈利丰厚，但无助于从根本上解决全行业关键基础设施瘫痪的问题。 确定性AI系统与企业变革管理 (23:45) 在确立现代数据本体与真实记录系统后，Valon 能够引入针对复杂长尾场景的 AI 智能体编排。例如在自然灾害发生时，系统可动态调度智能体匹配各地合规要求，为借款人制定纾困方案，并支持小样本贷款群体的对照测试（Champion-Challenger）。与通用编程助手不同，房贷场景要求绝对的“确定性准确度”，一旦 AI 在托管赤字或止赎流程上产生幻觉将引发重大法律风险。因此工程重点在于向量数据库、严格上下文排序与模型微调，确保模型在证据不足时明确回答“不知道”。 随着 Rithm Capital / Newrez 宣布向 Valon 迁移 400 万笔贷款，Valon 在转变为纯软件厂商后的半年内签署了逾 2 亿美元商业合同，并采用与客户降本增效成果直接挂钩的定价机制。面对大型机构数千名员工的系统替换，Linda Du 指出，软件落地的本质是一场直面内部政治与利益分歧的变革管理。此外，Andrew Wang 认为“服务”（Servicing）作为资金流动、严格监管与后台运营的交汇中枢，其核心架构未来同样适用于商业地产及医疗收入周期管理（RCM）等高摩擦产业。 该访谈展现了 Valon 通过“先做受监管实体业务，后售核心软件”的迂回战略打破万亿美元级产业的旧系统垄断。然而其经验亦表明，在高度受控且深陷路径依赖的基础设施产业中，技术突破无法脱离沉重的牌照合规周期与复杂的人事变革管理而独立成立。

## Background 本期内容由 a16z 普通合伙人 Angela Strange 与 Valon 创始人 Andrew Wang、Linda Du 共同录制，于 2026 年 10 月 7 日发布于音频播客平台。节目深入探讨了 Valon 如何在依赖互联网前过时系统、规模高达 13 万亿美元的美国住房抵押贷款市场中，从底层重构现代化的关键基础设施与 AI 落地范式。 ## TL;DR Valon 放弃了向传统机构兜售外包 SaaS 的平坦路径，选择直面监管死锁，亲自成为持牌服务商，将数十年的联邦及各州法条代码化为底层本体论。凭借比传统机构高出三倍的运营效率和 70% 至 80% 的运营利润率，Valon 在自营规模突破 2000 亿美元 UPV 后果断将服务实体出售给 Carrington，彻底转型为纯软件平台 Valon OS。公司通过打造高置信度的 Agent 军团实现房贷复杂长尾流程自动化，并以结果导向的商业模式迅速斩获 NewRez 转移 400 万笔贷款等数亿美元大单。团队更进一步揭示了“万物皆服务”的底层逻辑，强调在万亿级传统金融组织中落地现代 AI 基础设施，本质上是一场深刻的组织变革管理挑战。 ## Atmosphere & style 对谈整体洋溢着一种务实、极客且极其笃定的硬核工程气息。Andrew Wang 与 Linda Du 毫无空谈概念的浮躁，而是以坦率甚至带着自嘲的幽默，生动复盘了在政府邮务室人肉寻信、背诵法规条号以及硬抗复杂对账系统的艰辛往事。随着对话从底层牌照攻坚自然过渡到 AI 智能体编排与万亿级企业变革管理，整场交流展现出一种将智力全然倾注于修补现实破损世界的理想主义与极高认知密度。 ## Bulletin - `00:00` 全美规模达 13 万亿美元的住房抵押贷款市场长期被前互联网时代的遗留系统垄断，Valon 通过全栈自建持牌服务商打破僵局，其自研系统运营效率达到传统机构的 3 倍，运营利润率达 70% 至 80%。 - `20:00` 在自营贷款本金余额规模达到约 2000 亿美元后，Valon 依既定路线图将自营服务商业务出售给 Carrington，全面转型为对外输出纯软件平台的 Valon OS。 - `35:00` 转变为纯软件平台后的 6 个月内，Valon 签约合同金额突破 2 亿美元，其中包括推动大型房贷机构 NewRez 迁移 400 万笔存量贷款的历史级大单。 - `15:35` Valon 将联邦与 50 州法律编码为统一的数据本体架构，并以此为底层支撑构建出可动态编排的 Agent 军团，用于自动化处理灾害应对和长尾复杂工作流。 ## Key takeaways - `00:00` 美国 13 万亿美元的房贷服务底层由前互联网时代的单一旧系统垄断，且实际运作中资金对账工程缺陷极其严重。 - `05:00` 传统核心系统因数据建模能力缺失仅记录静态切片，一旦遭遇借款人去世或保险激增等突发状况，便会引发连环业务死锁与客户焦虑。 - `10:00` 面对纯软件 SaaS、并购整合 Roll-up 与从零自建全栈服务商三种路线，Valon 选择了最艰难但架构最彻底的自建持牌路径。 - `12:35` 为了跨越监管机构“已盈利”等先有鸡还是先有蛋的资质死锁，Valon 团队历时整整 3 年才拿下纽约州牌照，甚至曾亲赴监管机构邮务室找回申请文件。 - `15:35` 将海量联邦法规（RESPA、TILA、FDCPA 等）与 50 州地方条文提炼并系统性代码重构为底层数据本体，构成了 Valon 最坚固的合规护城河。 - `18:25` 向保守的资产管理机构推广时，Valon 凭借自研系统降低的成本让利给客户，并通过承受多年各级实地合规审计构建了“安全与紧迫”的双轮模型。 - `20:00` 自营房贷服务在达到约 2000 亿美元 UPV 规模后被果断剥离出售，以避免技术红利被禁锢在自有服务商体内，全面聚焦以 Valon OS 解决行业核心基础设施问题。 - `23:40` 依靠统一的数据本体支持，Valon OS 能够通过编排 Agent 军团动态处理灾害应急联络等复杂任务，打破了传统模式的同质化僵局。 - `25:00` 面向复杂长尾场景与人机交互，Valon 引入对话式 Harness 与置信度严苛约束机制，宁可回答“不知道”也杜绝模型幻觉产生。 - `30:00` “万物皆服务”：贷款服务基础设施具备极强的横向延展性，能够将底层能力复制到商业地产甚至医疗领域的收入周期管理（RCM）。 - `35:00` NewRez 确认迁移 400 万笔贷款，Valon 摒弃按坐席收费的传统模式，将商业收益与客户省钱或创收的实际结果完全绑定。 - `35:00` 大型受监管机构的软件落地本质是一场组织变革管理（Change Management），需要团队兼具顶尖技术拆解能力与复杂的组织沟通协调能力。 ## Detailed walkthrough ### `00:00` 13万亿美元巨头林立的未颠覆市场与创业起因 在现存所有未被数字化颠覆的传统行业中，住房抵押贷款（Mortgage）绝对稳居前三名。这是一个体量高达 13 万亿美元的庞大消费者债务市场，然而支撑整个行业运转的底层核心系统，绝大多数却垄断在一家建立于互联网诞生之前的传统老牌软件公司手中。这并非像其他成熟行业那样、由新创公司与现代软件巨头争夺最后 20% 的边缘市场，而是整个行业甚至连 2000 年代的现代技术标准都未曾达到。这代表着一块巨大的市场空白。 对 Valon 联合创始人 Andrew Wang 而言，进入这个领域的契机充满戏剧性。他在进入金融行业前毫无金融学背景，从未上过一门经济学课，甚至几乎搞不懂货币时间价值（Time Value of Money）。他最初的兴趣完全纯粹于数学，而房贷领域恰恰大量运用了随机微积分（Stochastic Calculus）来做提前还款模型（Prepayment Models）。但真正深入行业后，他发现现实中的各种金融模型大多只是“竖起手指测风向”，凭空胡猜还款速度。相反，房贷服务（Mortgage Servicing）背后庞大的数据量、资金流转与对账逻辑，暴露出极其致命的系统工程缺陷。对账系统在运作中资金账目几乎从未真正咬合匹配过，内部人员需要花费海量精力去人工修复这一漏洞。 这段痛苦的系统对账经历促使 Andrew 思考：必须重构更优秀的底层基础设施。他经历了从数学到系统、再由系统痛点催生创业的演变过程。对 Valon 来说，部署这类系统属于最艰难的企业级（Enterprise）替换：不仅要在极其严苛的监管环境中将旧核心操作系统连根拔起（Rip and replace），还要对数千名员工进行培训。这也是 Valon 团队锤炼应用 AI 能力的绝佳试验场，正如 Linda Du 所指出的那样：“未来十年市场所渴望的人才技能，就是应用 AI 的能力。而我们的团队正通过在最艰难的行业里实战来掌握它。” - **房贷服务核心定义**：它不仅处理资金流转，更包含大量法规遵从与高度复杂的后台运营，其性质与医院的收入周期管理（RCM）类似，本质皆为 Servicing。 - **垄断现状**：全美 13 万亿美元的房贷资产主要跑在一套诞生于前互联网时代的单一旧系统上。 - **数学与工程断层**：复杂的随机微积分模型在实际执行中多依赖主观猜测，而底层的数据流转与对账工程却存在严重系统缺陷。 --- ### `05:00` 架构陈旧对普通房贷客户的真实伤害 对于每一个按时还款的普通借款人而言，基础设施的优劣或许感知不深。但大多数美国家庭的财务缓冲极为脆弱，往往只差一张薪水支票（One paycheck away）就会面临违约风险；当突发变故来临时，底层基础设施的落后便会直接转化为房主巨大的心理压力与焦虑。 ```mermaid flowchart TD accTitle: 传统旧系统与现代化系统在突发事件中的数据流对比 accDescr: 对比传统房贷系统因数据模型缺失导致的断层与现代系统基于完整上下文的自动处理机制。 subgraph Legacy["传统旧核心系统 (1960年代架构)"] L1["借款人变更 / 房主去世"] --> L2["当前切片覆盖旧记录\n历史丢失"] L2 --> L3["存入外部临时数据库\n(Ad-hoc Database)"] L3 --> L4["拒绝家属调阅 / 人工反复核查\n造成客户严重精神内耗"] end subgraph Modern["Valon 现代化架构"] M1["借款人变更 / 保险激增"] --> M2["保留贷款全生命周期上下文\n(高保真历史追踪)"] M2 --> M3["规则引擎自动匹配监管条款\n(例如 36个月摊销托管差额)"] M3 --> M4["即时自动化审批\n无需上报主管人工流转"] end ``` 这种糟糕体验的根源在于旧系统的底层数据建模能力不足。传统系统只记录当前时间点的静态切片（Snapshot），无法高保真地保留一笔房贷全生命周期的完整历史上下文（Full context）。例如，当原有借款人去世、其子女接管房产处理遗产税时，旧系统会直接切断历史记录并回应“你不是原借款人，无法查看历史”。家属不得不被迫在承受丧亲之痛的同时，一遍又一遍地向客服重复解释家庭悲剧以索取记录。即便引入前沿的 AI，如果底层的核心数据模型本身没有正确表达这些信息，AI 也无法凭空产生正确的系统输出。 在更常见的场景中（如借款人无法按时还款或面临托管账户激增），旧技术的限制同样显露无疑： - **托管账户（Escrow）压力**：房贷托管账户通常按月预提用于支付年度房产税和房屋保险。近期美国部分地区火灾等自然灾害频发，导致保险费用激增，许多房主无法一次性承受每月剧增的账单。 - **僵化的摊销计算**：虽然相关法规与投资者准则允许将短缺金额分摊到 36 个月还清，但 1960 年代构想的旧系统往往死板地锁死在 12 个月内计算，将合理的分期方案变成必须上报主管的人工特批流程（Ad hoc approval process），耗时耗力。 - **数据模型的历史债务**：正如 Linda Du 强调：“当这些服务商的技术系统被架构出来时，那是 1960 年代。因此，即便是像财产保险、房屋抵押贷款保险这样基础的事物，都在随着时间不断演进。底层数据模型根本无法反映当今世界的运转方式。”当年为了适配落后系统而强行塞入的畸形数据结构，往往不会在第一周暴发，而是在五年后房主试图偿还托管账户差额时引发连环错误并不断滚雪球。 ### `10:00` 破局三大路径与从零自建的选择 房贷服务中，底层系统与用户体验紧密捆绑：即便拥有世界上最富同理心的客服代表，若核心记录系统（System of Record）计算扣款发生错误，也根本无法从根本上修复屋主的体验。面对这个技术严重落后、由老牌巨头主导且体量高达 13 万亿美元的庞大市场，团队在创业初期曾审视过三种可能的破局方式： 1. **纯软件模式（SaaS）**：仅开发底层软件，向老牌贷款服务商（Servicers）兜售，类似于法律科技领域 Harvey 的模式。但这在房贷领域行不通，因为行业监管极严，没有任何服务商敢贸然采用全新的从零开发系统（De novo system）；即便勉强签下一家标杆大客户，软件的研发方向也会彻底受制于客户既有的做事习惯，最终很可能只是“花了十年时间给老旧系统套上了一层漂亮的 UI 外壳”。 2. **并购整合模式（Roll-up）**：收购数家传统服务商，再向其内部注入现代化技术。这种模块化逐步替换旧组件的方式看似方便，但在长期架构上是错误的，并且还会深陷现有老旧平台的知识产权（IP）纠纷泥潭。 3. **从零自建全栈服务商**：自己成为持牌服务商，从第一笔贷款开始从零手写所有底层代码，让软件伴随实际业务规模共同成长。 Valon 最终选择了最痛苦、耗时最长，但战略上最彻底的第三条道路。Andrew Wang 强调说：“我们走的就是最后一条路，因为它是最痛苦的，耗时也最长。但第一种只是做传统软件公司的模式在房贷领域根本走不通，因为监管太严了，根本没有服务商会用一套全新的系统。” ```mermaid flowchart TD accTitle: Valon 破局路径评估与选择 accDescr: 展示 Valon 评估的三种商业模式以及为何最终选择从零自建房贷服务商与软件底层。 subgraph Evaluation["破局路径权衡"] O1["路径一：纯软件（SaaS）<br/>卖给老牌服务商"] -->|"痛点"| R1["监管严苛无人敢试<br/>沦为定制化并重造旧 UI"] O2["路径二：并购整合（Roll-up）<br/>收购传统服务商并替换模块"] -->|"痛点"| R2["知识产权 IP 纠纷<br/>架构被既有模块绑架妥协"] O3["路径三：全栈自建<br/>亲自拿牌照并从零写系统"] -->|"优势"| R3["享有最高架构自由度<br/>从 1 笔贷款平稳扩展至近百万笔"] end R3 --> Choice["最终执行路径：端到端从零构建服务商"] ``` --- ### `12:35` 牌照申领的“死锁困局”与物理邮务室寻信 自建全栈服务商的第一道难关是牌照。外界常误以为办牌照无非是填表递送，但实际上面临着极为荒谬的“先有鸡还是先有蛋”的监管死锁。例如，纽约州监管机构以及 Fannie Mae（房利美）、Freddie Mac（房地美）等机构通常有一条硬性前置要求：申请企业在拿牌前必须已经盈利。对于一家初创软件公司而言，这几乎构成不可跨越的门槛。 为此，Valon 必须极度严谨地倒推每张牌照的获取顺序。由于拿不到全国性政府贷款资质，团队决定先从非合格抵押贷款（Non-QM loans，仅需各州牌照）切入。然而这在房贷大州加利福尼亚又遭遇了阻碍：加州要求服务商必须先拥有政府贷款牌照，除非公司持有加州房地产局（Department of Real Estate）的经纪资质。Valon 只能找到具备房地产经纪资质的合伙人共同操作。更严苛的是，监管部门还要求团队在催收、客服、付款处理、逾期管理和止赎（Foreclosure）等房贷服务的每一个独立环节中，全员必须拿出 5 年以上的从业经验证明。 这些连环审批相互堆叠、互为前置条件，行业内按最优化步骤推算通常至少需要 3 到 5 年。在处理极其严苛的纽约州牌照时，甚至遭遇了更戏剧性的事件： - 团队最终历时整整 3 年才创纪录地拿到了纽约州房贷服务牌照，但即便如此，至今依然未获准在纽约州进行房贷发起（Origination）业务。 - 纽约州监管机构在审核时一旦漏审或驳回，时钟就会直接被重置 3 到 6 个月。 - 就在新冠疫情暴发前夕，监管机构告知申请文件“丢失了”，眼看又要延误半年。Andrew Wang 凭借快递单号确认文件已入库，索性直接亲自杀到了监管机构的办公大楼邮务室，翻找并当场找回了那份申请包裹。 --- ### `15:35` 重构法律法规：把海量法条沉淀为底层架构 办下牌照后，第二座大山是将数十年层层演变的联邦和各州法规完整翻译并编码进软件底层。Angela Strange 将这一漫长痛苦的过程戏称为“终极痛楚（Pain and suffering）构成的最大护城河”。 ```mermaid flowchart LR accTitle: 法律代码化转换流程 accDescr: 描述 Valon 如何将复杂的各州与联邦金融法规系统性重构成软件底层架构与业务本体。 subgraph Laws["法规梳理"] Fed["联邦法规：<br/>RESPA / TILA / FDCPA / GLBA / TCPA"] State["50 州地方具体法规：<br/>催收、止赎、隐私、第三方托管"] end subgraph Engineering["工程架构重构"] Refactor["抽象法律法条共性<br/>对整个法律逻辑库进行系统重构"] Schema["构建统一架构图式 (Schematics)<br/>建立清晰业务本体 (Ontology)"] end subgraph Verification["实际运行验证"] Audit["历经 5-6 年真实各州审计<br/>Fannie / Freddie / Ginnie 严苛合规检验"] end Laws --> Refactor Refactor --> Schema Schema --> Audit ``` 团队逐字阅读并人工标注了所有法条，包括覆盖房地产结算程序的 RESPA、借贷真实性法案 TILA、公平债务催收法案 FDCPA、金融服务现代化法案 GLBA、电话消费者保护法案 TCPA，以及各州关于止赎、托管账户和隐私权的复杂条文。因为许多州的法条演进自其他州，团队对庞杂的法理体系做了一次大规模的“法律代码重构”，提炼出能够通配各州差异的抽象架构。 在新冠疫情居家隔离的几个月里，技术团队埋头研读法规、撰写模式架构草案（Schematics），奠定了整套系统的基础本体论（Ontology）。正如 Andrew Wang 随后所调侃的那样，这期间的沉浸钻研让他在社交聚会上甚至能张口背出各州的具体法规条号，成为一种奇特的技术“极客派对绝技”。 - 覆盖全流程领域：包括催收规范、止赎法律程序、隐私权保护政策、托管资金（Escrow）核算机制等。 - 经过后续 5 到 6 年的连续真实验证与压力测试，这套系统才最终确保了业务在各个极端法律边缘情景下严丝合缝地运作。 --- ### `18:25` 资产管理人的经济算盘与企业级软件的“安全与紧迫”法则 当平台最终搭建就绪，面临的第三道难关是如何说服极其保守的资金方和房贷机构，将动辄数亿美元的高价值受监管资产托付给一家新公司。团队发现，推动房贷资产管理人（Asset Managers）尝试新事物，靠的不是浮夸的技术概念，而是纯粹的经济利益计算。 软件本身所带来的极高效率与优异的单量经济效益（Unit economics）直接转化成了产品壁垒。底层自研平台大幅降低了运营成本，Valon 选择把节省出来的利润空间让渡给资产客户。正如 Andrew Wang 所剖析的：“拥有这些房贷资产的资产管理者全都是纯粹的经济动物。说到底，只要你把价格降下来，就能促使市场屈服让步；而能做到这一点，前提是底层技术有能力支撑这样的成本结构。” 而在后来将这套系统作为核心软件（SaaS）推向金融机构客户时，Valon 内部总结出了一套双轮平衡模型： - **安全性（Safety）**：金融企业首先关注风险控制。正因为 Valon 自营着真实的服务商业务，承担了全部运营风险，并平稳度过了过去 6 到 7 年中所有严格的州级现场审计、Fannie Mae、Freddie Mac 以及 Ginnie Mae 的多轮合规审查，才为外部客户提供了坚如磐石的安全背书。 - **紧迫感（Urgency）**：在守住合规底线的基础上，再通过技术创新所带来的利润率差异和自动化潜能，在行业中创造出令传统机构无法忽视的转型紧迫感。 ### `20:00` 飞轮效应与战略取舍：从自营服务商到纯软件操作系统 启动任何具有复利效应的事物总是最艰难的，核心在于如何从零建立并积蓄势头。Valon 破局的第一步是亲自下场建立贷款服务机构（Servicer），直接在市场上与传统玩家同台竞技。这种策略直接催生了紧迫感，因为 Valon 实际上是在与自己未来的潜在软件客户直接抢夺市场份额，寄希望于日后再赢得他们的理解与合作。在传统房贷服务这个利润微薄、近乎保本的行业中，Valon 凭借自主构建的技术架构展现出了惊人的经济效益差距：其运营效率大约是传统同行的 3 倍，直接将这门盈亏平衡的生意转变为拥有 70% 至 80% 运营利润率（operating margin）的高利润业务。 这种断层式的利润率优势构成了极强的说服力与行业紧迫感。当 Valon 拿下第一个大型客户后，飞轮效应迅速启动，其他五家头部巨头便主动找上门来探听虚实。截至出售前，Valon 的在管贷款本金余额（UPV）已达到约 2000 亿美元。然而，他们并没有选择顺水推舟继续做大规模，而是宣布将旗下房贷服务业务 Valon Mortgage 出售给 Carrington，全面转向对外输出 Valon OS 纯软件系统。这一步并非临时起意，而是早在最初融资时就写进了一份长达 60 页的 Series A 商业计划书 PDF 中，团队自始至终都在按部就班地执行这套“主宰市场的蓝图”。内部曾多次探讨过是否要继续做服务商——毕竟市场上存在估值超百亿美元的纯服务商企业，做大规模是一条现成且行之有效的路径。但正如 Andrew Wang 指出，每一次讨论的终点都高度一致： > “The thing is, is that if you keep it as a servicer, you're keeping all of the technology and the alpha for yourself. And at the end of the day, you didn't actually change anything. You didn't fix the core infrastructure problem.” 如果仅仅停留在做一家高效的服务商，所有技术红利与超额收益（alpha）都会被局限在自家围墙之内，底层的核心基础设施问题并未得到真正解决。公司的使命绝不仅仅是多赚钱，而是要改变行业、修复这个破损的世界，带领大家“从大型主机（mainframes）、陈旧遗留软件和破碎的关键基础设施中拯救世界”。 - **运营利润率对比**：传统房贷服务业处于盈亏平衡边缘（接近 0% 利润率），Valon 达到 70% 至 80% 的运营利润率。 - **业务规模**：自营服务规模达到约 2000 亿美元 UPV 后果断脱手，专注纯软件生态。 - **战略交易**：将房贷服务实体剥离并出售给 Carrington，全力将 Valon OS 赋能给整个行业。 --- ### `23:40` 生成式 AI 赋能：从确定性脚本到“Agent 军团”编排 Valon 创立之初，生成式 AI（Generative AI）尚未成熟。但随着底层大模型与企业级基础设施的演进，房贷服务展现出了极其完美的契合度。在过去的传统模式下，后台运营人员必须应对极其繁琐的合规与处置流程。以灾害应对为例，当某地发生自然灾害时，服务商必须立刻致电业主，确认其人身与房屋安全、了解其后续计划、排查是否需要紧急援助等。虽然常规流程都备有标准话术，但要根据当地政策、灾害级别或个性化需求临时动态调整通知策略或处置方案，在传统体系下几乎是一项不可能完成的任务。 传统机构唯一的应对方式就是打印出一叠又一叠的电子表格，将不同批次的贷款清单分发给各个员工小组，机械地交代某组做 A、某组做 B，全靠人工去摸索执行。而在当今 Valon 的产品套件中，运营人员不再是单兵作战的操作工，而是掌控着一支“小型智能体军团”（mini armies of agents）。系统可以瞬间编排一整套 Agent 自动发起外呼、联络业主、动态提供差异化的救济方案并执行端到端的工作流。这种能力不仅大幅提升了运营效率，更赋予了服务商前所未有的可定制性，打破了过去因零利润率而不得不采用“千人一面”（cookie cutter）的同质化服务模式，让房贷服务从大宗商品化彻底走向专业化与特色化。 ```mermaid flowchart TD accTitle: 传统服务商与 Valon OS 在灾害应急处理上的架构对比 accDescr: 对比传统手动打印电子表格、人工打电话的流程，与 Valon OS 利用数据本体驱动 Agent 编排自动执行的端到端流程。 subgraph Legacy["传统服务商模式 (低效且僵化)"] L1["突发灾害事件"] --> L2["打印电子表格清单"] L2 --> L3["分发清单给人工呼叫团队"] L3 --> L4["人工对照固定脚本逐一外呼"] L4 --> L5["无法根据细分场景动态调整"] end subgraph Valon["Valon OS 智能化工作流"] V1["突发灾害事件"] --> V2["底层结构化数据本体 (System of Record)"] V2 --> V3["编排 Agent 军团 (Mini Armies of Agents)"] V3 --> V4["动态适配各地区策略与通知方案"] V4 --> V5["自动化外呼 / 智能交互 / 救济方案推进"] end ``` --- ### `25:00` 复杂长尾自动化与人机交互形态的重塑 在 AI 爆发前，Valon 搭建了极为严谨的数据本体（data ontology），能够依靠结构化数据对高确定性的流程进行自动化，但能自动化的始终只是极小一部分规则明确的工作。如今，Agent 不仅能承担托管分析（escrow analysis）等繁复日常，更能覆盖原本属于人工的长尾与复杂场景。运营者甚至可以针对 100 笔贷款直接开展“冠军/挑战者”（champion challenger）式的 A/B 测试策略实验，尝试不同的通知策略与处置节奏，快速观察哪种方案能带来更优的资产组合表现。这种高度灵活的策略迭代在传统依赖人工严密培训的模式下是无法想象的。 与此同时，人机交互的形态也发生了根本性变革。过去，软件企业全力打造核心记录系统（system of record），人类通过固定的 UI 界面与之交互，企业甚至需要专门招聘熟悉特定界面跳转的熟练操作工；而在当今的 Valon OS 之上，系统在底层记录系统之上封装了专属 Harness，后台员工可以直接通过对话界面完成复杂的长尾任务。面向消费者（C 端业主），交互形式同样扩展到了对话式 AI 与语音 AI（voice AI）。在底层架构研发中，由于房贷行业对准确性（correctness）有着极致的严苛要求，Valon 针对行业特性深度定制了模型对齐与 Harness 机制： > “And so you would probably want to align the models as much as possible and really the overall harness outcomes so that if you don't know the answer, it generally says, I don't know versus let me give you some hallucinated answer.” 为了根绝模型幻觉，系统在检索与推理中被赋予了极高的置信度约束，在信息不明确时宁可直接回答“不知道”，也坚决杜绝编造内容。此外，相比编程等受限且封闭的场景，房贷服务涵盖跨周期的复杂生态系统与多方动态变化，Valon 正在持续构建能够利用真实交互数据进行持续学习（continual learning）的基础设施，确保整个智能化体系在瞬息万变的企业环境中保持精准与稳定。 ### `30:00` “万物皆可服务”：底层基础设施与跨行业扩张潜力 在建立起能够持续运行模型评估（evals）、抽取数据并根据不同客户的个性化需求进行持续改进的底层系统后，Valon 的野心远不止于住宅抵押贷款服务（Residential Mortgage Servicing）。Andrew Wang 揭示了一个鲜为人知的行业本质：贷款服务（Servicing）实际上是所有高度受监管企业的核心枢纽（nexus）。大众通常认为服务业务仅限于住宅抵押贷款，住宅房贷固然体量庞大、高度受监管且承载着美国人实现居者有其屋的梦想，但只要向周边迈出一步，就会发现商业地产（Commercial Mortgage）中潜藏着所有租户的财务信息和房地产运营数据。这类似于 Toast 等平台利用餐饮系统切入中小企业融资，掌握了商业抵押贷款服务，就天然掌握了更广泛的企业融资与数据入口。 更进一步看，医疗领域的收入周期管理（Revenue Cycle Management, RCM）本质上也是一种针对医院账单和患者赔付的服务系统，处理的数据实质上就是电子病历（EMR）。由于贷款服务业务具有极高的业务广度和复杂的法规门槛，一旦攻克这一难关，企业便能像触手一样延伸至业务的各个脉络。在住宅领域，传统服务商仅仅停留在催收月供上，而 Valon 结合自研软件与 AI，能够做到每月主动联络房主、了解其实时动态并建立长期信任。正如 Andrew Wang 所言：“万物皆服务（Everything is servicing）。”Linda Du 总结道，贷款服务正是支撑所有涉及资金流动、严格监管和复杂运营流程的底层关键基础设施，也是客户黏性最高、最复杂艰难的一环。 ```mermaid flowchart TD accTitle: 贷款服务业务跨行业延展架构 accDescr: 展示 Valon 从住宅贷款底层基础设施向商业贷款及医疗收入周期管理等领域的辐射路径 subgraph CORE["底层关键基础设施 (Valon 核心架构)"] direction TB A["底层系统架构\n(资金流动 + 强监管合规 + 运营工作流)"] B["AI 持续评估与模型优化引擎"] A <--> B end subgraph SECTORS["多元高壁垒垂直领域"] direction TB C["住宅房贷服务 (Resi)\n深度房主关系、月度主动触达"] D["商业地产服务 (Commercial)\n掌握租户财务与数据、企业融资"] E["医疗账单管理 (RCM)\n对接电子病历 (EMR)、医院理赔服务"] end CORE --> C CORE --> D CORE --> E ``` ### `35:00` 2 亿美元大单与 NewRez 转移 400 万笔贷款背后的共赢机制 当 Valon 验证了作为独立服务商的能力后，正式对外剥离服务业务并全面转型为软件平台公司。随后便迎来了行业震动性的里程碑：Rithm Capital 旗下的大型非银房贷机构 NewRez 决定将其 400 万笔贷款转移至 Valon 平台，成为行业历史上最大规模的贷款服务迁移之一。在激烈的竞标中击败所有老牌系统提供商的关键，在于 NewRez 此前通过 Valon 提供的子服务（subservicing）业务，已经在无风险的实操环境中亲眼见证了软件运行的极致性能。Andrew Wang 强调，NewRez 在过去合作中看到的效果甚至只是软件性能的下限。 更为决定性的因素在于商业模式与客户利益的高度绑定。在正式以软件公司身份推向市场的短短 6 个月内，Valon 就签下了超过 2 亿美元的合同。Valon 摒弃了传统软件商按坐席或固定服务收费的脱节模式，将自身收入完全锚定在业务成果上。正如 Andrew Wang 明确阐述的核心原则：“我认为一旦你失去了那种利益一致性，公司就开始走下坡路了。因此对我们而言，我们把一切都构造成以结果为导向：客户每赚一美元，或者客户每省一美元，我们才获得相应的收益。”这种利益一致性打消了客户的所有顾虑，让大型金融机构迅速下定合作决心。 - **商业拓展规模**：作为软件平台推向市场的前 6 个月内，签约交易金额突破 2 亿美元。 - **历史级迁移规模**：Rithm / NewRez 确认将 400 万笔存量贷款服务转移至 Valon 平台。 - **价值分成模式**：客户收费直接与节省的成本或创造的新增收益挂钩，实现结果导向。 ### `35:00` 人才画像、五年留存率与千万级组织的变革管理（Change Management） 谈及 Valon 位于纽约办公室的团队氛围与文化，Andrew Wang 观察到团队成员身上有一种鲜明的特质：他们拒绝在象牙塔中空谈，而是渴望亲自动手解决实际问题。Valon 吸引了大量来自顶级大厂以及 OpenAI、Anthropic 等领先 AI 实验室的顶尖技术人才。这些人清晰地意识到，底层模型训练和算力层面的问题已有无数聪明人在推进，而真正有价值、能够带来真实世界冲击的，是如何攻克这个高度复杂、严重依赖传统流程的巨额支柱产业。这种将智力投入到具有明确现实意义的硬核攻坚中的成就感，形成了极强的凝聚力，使得公司管理层中有 75% 至 80% 的人员在 Valon 任职超过五年。 在向超大型企业客户部署平台的过程中，挑战的本质也发生了转变。Linda Du 直言，如果在六年前，她会认为部署只是一道纯粹的技术题；但如今经历过真实的落地过程后，她认识到这本质上是一场组织变革管理（Change Management）。万人规模的大型金融机构中，不同层级的员工各自拥有不同的目标函数，推动变革极其困难。Linda Du 指出：“你花了过去 6 到 12 个月建立起来的正是那种变革管理的肌肉记忆，即你如何穿梭于复杂的组织之中？你如何促成变革发生？事实证明，在规模化体系中推动变革是极其艰难的。但我认为，在我们当前所处的 AI 时代，这正是未来十年最大的价值驱动力。”能够在这一领域脱颖而出的复合型人才，不仅需要卓越的技术问题拆解能力，还必须具备深刻的人际洞察与沟通协调能力，唯有将全局目标与个体诉求调和统一，才能让前沿软件真正在万亿级市场中扎根运转。 ## Notable quotes > `18:25` 只要你把价格降下来，就能促使市场屈服让步；而能做到这一点，前提是底层技术有能力支撑这样的成本结构。 — Andrew Wang > `35:00` 在规模化体系中推动变革是极其艰难的。但我认为，在我们当前所处的 AI 时代，这正是未来十年最大的价值驱动力。 — Linda Du > `00:00` 未来十年市场所渴望的人才技能，就是应用 AI 的能力。而我们的团队正通过在最艰难的行业里实战来掌握它。 — Linda Du > `35:00` 我们把一切都构造成以结果为导向：客户每赚一美元，或者客户每省一美元，我们才获得相应的收益。 — Andrew Wang ## Mentioned Angela Strange, Andrew Wang, Linda Du, Valon, Valon OS, Valon Mortgage, Carrington, Rithm Capital, NewRez, Fannie Mae, Freddie Mac, Ginnie Mae, Toast, OpenAI, Anthropic, Harvey ## Video description a16z General Partner Angela Strange sits down with Valon’s Andrew Wang and Linda Du to unpack what it takes to rebuild the infrastructure underneath a $13 trillion mortgage market that still relies heavily on systems designed before the internet. Linda and Andrew explain why Valon chose the hardest path: becoming a regulated mortgage servicer, translating decades of federal and state regulation into software, and proving the platform on its own loans before selling it to the industry. That foundation made Valon roughly three times as efficient as traditional servicing and created the system of record it is now using to bring AI into complex mortgage workflows. They also discuss what AI makes possible on top of that infrastructure, from agents handling long-tail servicing tasks to voice interfaces and more personalized customer experiences. And they explain why deploying the technology into large regulated enterprises is ultimately as much a change-management challenge as a technical one. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/b653e837-66de-4c84-8582-adba2b75051f/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=b653e837-66de-4c84-8582-adba2b75051f&feed=JGE3yC0V)

[18:25] Andrew Wang: 只要你把价格降下来，就能促使市场屈服让步；而能做到这一点，前提是底层技术有能力支撑这样的成本结构。 [35:00] Linda Du: 在规模化体系中推动变革是极其艰难的。但我认为，在我们当前所处的 AI 时代，这正是未来十年最大的价值驱动力。 [0:00] Linda Du: 未来十年市场所渴望的人才技能，就是应用 AI 的能力。而我们的团队正通过在最艰难的行业里实战来掌握它。 [35:00] Andrew Wang: 我们把一切都构造成以结果为导向：客户每赚一美元，或者客户每省一美元，我们才获得相应的收益。

[5:00] Andrew Wang: One of the key problems with some of the legacy systems is that they don't really model and really capture all of the context associated with the life of the mortgage. [3:40] Linda Du: The skill set that people will be hiring for in the next decade is the ability to apply AI.

[0:00] Mortgage is top three in terms of undisrupted industries that exist. This is an industry that hasn't even caught up to 2000s technology yet. That's sort of the size of the opportunity in front of us. It's $13 trillion of consumer debt that basically runs on a single incumbent that built their legacy system before the internet was invented. It turns out revenue cycle management is just servicing for hospitals. Everything is servicing. It's the underlying critical infrastructure that supports basically anything that has money movements, some sort of regulation, and then an operational component. You started the company before Generative AI was really working. pitch deck. What is possible today is... Mortgage servicing is a $13 trillion market that still runs in large part on technology designed before the internet. Valen decided to rebuild it from scratch. In this episode, A16Z general partner Angela Strange sits down with Valen's Linda Du and Andrew Wang to unpack what it took to replace decades of legacy infrastructure in one of the most regulated parts of financial services. They discussed why Valen became a mortgage servicer before becoming a software company. how the team turned decades of regulation into code, and why operating the technology themselves was critical to proving a good work at scale. And now, AI is changing what can be built on top. Linda and Andrew explain how agents can handle workflows that were previously too complex to automate. why the underlying system of records still matters, and what they've learned deploying new technology into enterprises with thousands of employees. you So Andrew, I think a great place to start would be from your time at Soros and how you decided that mortgage servicing was going to be a really interesting problem to work on. So a fun story is that when I got into the world of finance, I actually had zero background in finance. I never took a course in economics. I barely understood what the time value of money was. And so my background was going to be a big part of the world. general interest in finance was actually just in math. And what people don't really know about mortgages, which is just like a very random esoteric fact, is that it's actually a place where there's a lot of stochastic calculus. Prepayment models have stochastic calculus. And so when I started the job, I got very interested in making sure I mastered stochastic calculus. I had taken a course in college. I was going to do all that. And it turns out, it's all a finger in the air. Just guess your way through math. So I'm going to do that. That wasn't really, yes, yes. You just need to guess some number. You're like, yeah, I think it's going to prepay at X speed. Anyways, so that was my initial reason to get into mortgages. What happened thereafter was you end up realizing working with many of the mortgage servicers that it's a big systems engineering problem because there's a ton of data, a lot of money moving around, a lot of entries. They have to constantly be reconciled. affiliation systems to make sure the money tied. Turns out it never tied. And so I spent a lot of time trying to fix that problem internally. And that got me really interested in figuring out, hey, how can I actually build better infrastructure? How can I invest maybe in a company, do it with a company, and actually ultimately build better systems so that none of this would be a problem. So it really started originally off of systems and my own pain. Maybe I would say it started off with math. Math became systems, systems became pain, and here we are today. And so discovering a market has a lot of technical problems to solve is one piece of it. Maybe Linda, from a market perspective, describe the state of the world as a mortgage servicing industry before Valen came in. Yeah, the way that I think about it is mortgage is probably top three in terms of undisrupted industries that exist. And so it's $13 trillion of consumer debt that basically runs on, for the most part, a single incumbent that built their legacy system before the Internet was invented. So we're not talking about for most industries, you are fighting with modern incumbents, new startups for the last 20%. We're talking about the entire. It's like this is an industry that hasn't even caught up to 2000s technology yet. And so that's sort of the size of the opportunity in front of us. And from a deployment standpoint, these are like the biggest, baddest enterprise deployments to exist. You have to train thousands of employees. You have to basically rip and replace the core operating system. into the real world in a very heavily, heavily regulated industry. And what we always tell the team is that the skill set that people will be hiring for in the next decade is the ability to apply AI. And our team's learning that by doing it in the hardest industry possible. [5:00] Well, yeah. So I think big technical problems, very old industry. I think one of the other interesting things about this industry and one of the reasons I love investing in infrastructure is if you unpack crappy consumer or business experiences, it usually often comes down to the plumbing. And so maybe if I'm a homeowner and my home is being serviced by one of the incumbents versus Valen, what might I notice from just a human perspective? Yeah, it's a... Classic problem where the average person who pays on time probably doesn't have to notice that much. I mean, this is a very classic thing where I think whenever people realize that most Americans actually are barely maybe one paycheck away from being able to make their payments. It's a little bit of an astonishing fact, but it's a similar dynamic where in mortgages, again, most people pay on time, but there are people who... have certain events that happen to them that are unfortunate, not predictable, and really cause a lot of stress and anxiety. So here's an example of this. One of the key problems with some of the legacy systems is that they don't really model and really capture all of the context associated with the life of the mortgage. Meaning you originate the mortgage, they have a snapshot of what's on right now, but the history isn't fully with really enough fidelity captured. as maybe they were the borrower on the loan before, and then... Thereafter, there's a new borrower on the loan. That is a normal thing. The systems today do not actually capture all of the changes along the way. And so if you want to look at what happened before that, It's actually very challenging for the servicers because they put it in an ad hoc database on the side. They only have what is going on at any given point in time. And so you might have a situation, as an example, where someone says, hey, or, not says, something cursed them, they end up in a situation where a homeowner passes away, their kids are trying to take over, but they no longer have access to the records before. And that might be important because they're trying to file state taxes, they're trying to figure out what happened before. And because of how, again, the systems work, they'll say, well, you weren't that borrower, you don't have access to that history, we're cutting it off at this point in time. And so that causes a lot of anxiety and stress, because you have to keep in mind, right, these are like real people at a moment in time where they're trying to figure out something And you have to over and over again tell them, my mom passed away. This is why this happened. Please give me the information. And so it percolates in that form and fashion. AI obviously helps a lot of this because you can get that context. But it doesn't change the fact that the information must be represented in the right way so that the systems can perform as expected. And so really, the more information represented, the better way possible, the better the experience ends up being. of a customer if you helped. Even the more common, simpler case of, I can't make my payment this month is something that traditional mortgage servicers can struggle with. So how is that different in Valon? Or if it's kind of multiple times in a row, you might need a loan modification. Yeah, there's all sorts of variations of this. Again, one of them is There is this concept of the escrow when it comes to mortgages. It basically contains both taxes and insurance that need to be paid over time. They generally happen on an annual basis to make sure that homeowners are reserving for the right amounts. You sort of pay it on a monthly basis and you accrue a reserve to pay for it when the time comes. One of the problems right now that everyone faces is actually insurance costs and taxes are actually exploding. And so many homeowners aren't able to make that large payment or change a payment in such a quick fashion. That requires usually a pretty annoying process because one, most systems are designed so that they take the amount and they basically just calculate how much it costs over the course of a year. Even though the regulation and many investor standards allow you to amortize it over 36 months. And so that becomes a ad hoc approval process that has to go to a supervisor as opposed to having really great systems where you pretty much can immediately determine should they be approved, should they not be approved. It might be some big fire happened. And so many of the insurance companies are raising prices. Being able to offer that right experience, really automated and really give comfort to the homeowners in their time of need. That is really important for a servicer and that ultimately builds the best customer relationship. And I think it's hard for people to conceptualize how old the existing software is because you have to think about it like when these services, technology systems were architected. This was the 1960s. And so even something as basic as property insurance, property insurance, mortgage insurance keeps evolving over time. And so the data models fundamentally just don't model the world the way that it behaves today. And so you end up with sort of over a decade, something that someone forced into a system that didn't actually model it correctly. And it's not like the problem is going to happen that week. It's like the problem will happen five years later when someone's trying to make their escrow payment and it compounds. And so that's the part where it's like, [10:00] you can have the world's best customer experience team and the most empathetic call center agent in the world. But if fundamentally you were charged the wrong amount of money because the system of record was flawed, then you can't fix that homeowner experience without fixing the infrastructure. This is actually a perfect segue. I want to talk a little bit about how Valon got started. And the flip side of these technologically behind incumbent-driven large industries is they can be very hard to break into, right? And so, you know, we had three options. One, build the software and sell it into the incumbents. And that's been successful in many industries, right? Like Harvey and Law is doing that right now. The second would be do some kind of a roll up, buy a few of the incumbents and then inject them. And then the third, which is what you chose, which is, no, we're just going to build the entire thing from scratch. And so maybe talk through like why that option to get started. Yeah. So we went through all three options. I don't know. I'm sure. And just to be clear, the path that we went down was the last one because it is the most painful and it takes the longest. But really option one, which is just building, you know, a traditional normal software company it doesn work for mortgage because it so heavily regulated that no servicer is going to use a de novo system And then even I would say from a R perspective it like if you let say that we managed to get one big marquee customer to sign up, we would basically be beholden to how they want to do things, which will look an awful lot like how they do things today. And so you just run the risk that you wake up 10 years from now and you rebuild the legacy system with like a really pretty UI. So that crosses out number one. Number two, the problem is IP. So if you buy an existing servicer that's using an existing legacy platform, the nice thing about being able to build software, sorry, I should say convenient, it actually I think is not long-term correct, is you can basically swap things out module by module. But one, I think that actually forces you into platform design decisions that probably aren't right. And then two, you have this IP issue. And so really that left us with the choice of we got the privilege of starting our own servicer and then just building the software from scratch, not looking at anything. And the software got to evolve as the servicer was growing as well. So it's like we started with one loan, which is really nice, and then ended up with close to a million Yep And then well and I think that strategy has proven very prescient now But to get it started I think of it in three buckets One you needed to get licensed Two you needed to turn decades of regulation into code And then three, you needed to convince asset managers, lenders, to give this highly regulated asset over to a new company. So I think it's interesting to briefly touch on each of those. I think starting with licensing, I think it's easy to think, I fill out some long forms, I throw them over a wall, eventually I get a license. Maybe talk us through what was that process, which I think at the time was very painful and now is a very nice mode. Yeah, I would like to say we were so thoughtful that we already knew all these things going in. There was, kind of late, looking back, some slight improvements we could make in terms of the licensing process. But for the most part, all of those required. For context, mortgage servicing and many of these licenses have pretty much chicken and egg problems. One of the key requirements is if they tell you, you actually need to be profitable before you get a license. This is true for the state of New York and a couple of other regulatory approvals like Fannie and Freddie. They'll make some exceptions and you can get around it, but it's quite challenging. And so when you look at it from the get you end up in a situation where you first notice holy crap this is going to take a long long time And two then you end up in a situation where you have to really be methodical in plotting out what licenses you get and what order. So for us, what we did was we said, okay, we're going to start out with this type of customer. We need to figure out, because we can't get the major government licenses, we need to look for a specific type of loan that we can get licensed for. It's called at the time. They're still pretty prevalent today, and they're actually, I think, growing non-QM loans. They only require you to have state licenses. Here's the problem, though. With state licenses, you need to pick the states that people actually make non-QM loans. One of the biggest ones is the state of California. State of California actually has pretty stringent requirements. It actually requires you to have a government application or government loan approval in order to service California loans. There is one exclusion for the Department of Real Estate. if you're in the brokering business and a real estate brokering business. So now you need to find someone who is in the department or is a real estate broker to actually help you get that license. Simultaneously, you also need to show them that you have five years plus of experience across every part of servicing. That's collections, that's customer service, that's payment processing, that's delinquency, that's foreclosure. So you need to do all these different things just to get going. [15:00] And like, let's not even get started when it comes to, well, once you've done that, how do you actually make sure you get the subsequent approvals and everything else? But it's an extraordinarily long process. Most people forecast this type of endeavor taking three to five years at the minimum in order to, and that's if you take the most literal optimal steps, three to five years to get all of the different approvals because they all stack on top of each other. You can't do one without the other. And by the very end, you end up having to, really aggregate not just business, but really expertise across a very, very large set of domains. There's many years of stories, including you going to visit mailrooms. Exactly. Because I'll give you the classic one, which, by the way, is a funny one. We actually got the approval for New York in a record time of, I want to say, three years. So it was a record time of three years. And we actually never ended up getting approved to license or to originate mortgages in New York. That just tells you how challenging and how difficult it is. And they have this funny thing where apparently they will review your application, but if they either miss it or deny it or whatever else, they'll reset the clock like three to six months back. And this was right before COVID happened. And so I knew I wasn't going to be in New York. for the moment. And so I didn't want to wait another six months before I came back to New York. And so what I made sure I did was I went to the mailroom and I identified where the package was because they told me my application was missing. And I was like, I have tracking on this. I know it's there. Well, let me go visit your offices to actually track this down. Fantastic. All right. So we've made it through that full process. And then maybe talking about taking the regulation and putting it into a platform. I think it's really easy to knock on the incumbents, state regulations changed over time? How did we approach that from scratch? Thank you. Honestly, I think the... unfortunate reality, which is, again, like you've mentioned before, a little bit of a note, which is how we make ourselves feel better about all of it. The pain and suffering is just... Greatest note of all time, pain and suffering. But what we ended up doing was reading every single regulation. So there's a whole bunch of federal regulations. There's, you know, RESPA, there's TILA, there's FDCPA, there's GLBA, there's TCPA. There's all these like different regulations on the federal level. You have to cover their debt collection practices. You have to cover their real estate foreclosure practices. You have to cover their privacy practices, their escrow practices. So you have to cover every one of these. And you basically go through a very... I'm going to say it just as a fun exercise of refactoring the entire legal code base, if you can put it that way. It turns out many legal statutes are derived off of different states. And so you sit there basically figuring out how to create an abstract framework that handles all of it. And then you put it into your architecture and then you start to test it. But what really worked for us, quite honestly, was the fact that COVID happened. regulation and annotating and basically coming up with the schematics. And that's what got us through it. It was pure grit, effort, pain, suffering, all these different things. And then obviously, you have the five to six years thereafter of testing and making sure it works exactly as you sort of intended. Which we'll get into in a second, sets up a very nice ontology for where Valen is going next. There was a point at which you could cite almost every single state's regulation, and I can tell you what code number it was. All right. Good quiz. Yeah, we'll do a pop quiz. Exactly. Very fun at the dinner party. Amazing. And then, okay, so you've got it built, which is a huge challenge, but then convincing customers that probably were not leading in necessarily to new technology to start off with. on the system? Yeah. So I would say... You know, the framework that we use internally for software is... is a little different than how we started with the loan servicer. So with the servicer, the nice thing is that unit economics and efficiency kind of largely was the product, meaning that the better we made the technology, the more margin we had. And honestly, we just passed that margin through to the end customer. And all of these asset managers who own these mortgage assets, they are economic animals. And so at the end of the day, it's like when you lower price, it's like you can get the market to capitulate. And that just requires the technology From a software perspective, we think about it internally as it's really you always have to balance safety and urgency. And so you get safety because we built a servicer that actually used the technology and proved for the last, you know, six, seven years that it's gone through every state exam, every Fannie Mae audit, Freddie, Ginny, you know, all of the different sort of compliance hoops that you have to get through. And so it's safe because we actually took the operating risk of running our own servicer on the system. And then in terms of urgency, it's something [20:00] that you honestly Starting to compound something is always the hardest part, right? It's creating and generating momentum. And so for us, the first thing that we did is we took our servicer and we competed in the market. And so that's one way of creating urgency because now you're actually directly competing with your future software customers and hoping that they'll forgive you down the line. low margin business, Valen turned into a high margin business. Like what was the economic difference that everyone kind of perked up in the industry? Yeah. So we are about three times as efficient. And so you take this break-even business and you turn it into sort of a 70, 80% operating margin business. And that margin is what we used initially to create that urgency. And then it compounds, right? So it's like once you get your first big customer to sign up, then the other five big guys call you and say, hey, were you actually serious the flywheel effect. But without our own servicer in the beginning, it would have been really hard to generate that. Yep. And so you built Valen Servicing, I think in the 200 billion, how big? UPV, yep. Exactly. And then you could have kept scaling that. Yeah. But instead, you decided... You're going to sell the customers. We're going to go through software. Just to be clear, this was the original plan. It was what you were pitched at the beginning. I agree. I agree. I agree. This is actually one of the funny things that all the new employees who joined, they're like, wow. Like, You guys are selling this thing and say, this is always what we've been meaning to do. But we actually share our Series A deck with our... To be fair, you had an incredibly detailed Series A deck. I think it was like a 60-page PDF of the master takeover plan. And then we followed it step by step. You guys just recently announced the sale of Valen Mortgage over to Carrington. And now you're taking that OS and selling it into the industry. once in a while, as you would expect, right? Where we would sit down and we would say, hey, it is a fully viable path to just build a big service. There are $10 billion plus companies that do this. But for us, it always came back to every conversation ended with we actually want to change the industry. And the thing is, is that if you keep it as a servicer, you're keeping all of the technology and the alpha for yourself. And at the end of the day, you didn't actually change anything. You didn't fix the core infrastructure problem. Yeah. I think for companies who want to a clear purpose and mission. That's not just we're going to make more money. That's great. It's we want to change the world. We want to fix a rather broken world. It's extraordinarily important for the employees and really the direction for the company. Because you want people to not just dream, oh, I make enough money now. It's like, what can I do that's greater? What can I do that's bigger? What can I do to actually solve the impossible? We got to save the world from mainframes. Wait, but your mission statement is actually that. mainframes, legacy software. And broken critical infrastructure. Exactly. Maybe speaking of the bigger opportunity, you started the company before Generative AI was really working. What is possible today that wouldn't have been in your Series A pitch deck? How beautiful of a problem setup mortgage servicing is for what's about to happen, both with AI, but just overall enterprise and infrastructure and society. to say what is possible today is you can have these extraordinarily complex cases and you can actually research and basically provide the right set of information to the operators so that they're more and more in a world where they're effectively operating little mini armies of agents to actually solve whatever set of problems. So I'll give you a really simple example. When a disaster happens, you're supposed to make some phone calls, make sure that the homeowner is actually able to plans are, if they need any assistance, you have to go through all this process. And that is normal and that is fine. And people have different scripts to set that up. But changing that on the fly or handling very, very specific different notifications or tactics or plans per locale is... Basically impossible. The only way to do that today as a normal servicer is you basically print out a spreadsheet and you hand a list of loans to one group of people and hand another list of loans to another people. And you say, I want you to do A, I want you to do B, and I want you to go figure it out. Versus with our systems and really the power of what AI can do today, you could literally in our product suite say, hey, I want to orchestrate a set of agents to make phone calls, to contact homeowners, to provide them these different plans, run these different workflows. Um, scenario and [25:00] That, I think, provides a lot of Both. power to the underlying servicers and becoming more efficient, but it actually gives them a lot more customizability. And the latter is actually the really, really cool and important part, because it allows every servicer to be the best versions of themselves. They can really think about, "Hey, how do I want to handle a specific scenario that pops up?" Versus just saying, "Look, I need to run this quote unquote efficiently as much as possible running 0% margins. and I thus only have a cookie cutter way of going about things and I actually can't change it. It really turns the world from one that is really commoditized to one that is much more specialized and really flavor driven. Yeah. So there's one way to think about it. You've now created this very organized data ontology. And then pre-AI, any of the tasks you needed to do on top were definitely more efficient. Whereas now you can spin up, I think you've got dozens of these, different agents that can do like escrow analysis and sort of, Any human tasks that you might need to do on top of it so that packaged information. Yeah, it's a completely different world where before you could only automate things that were clearly deterministic because you could say, hey, I have the right structured data. I have the information. I can automate a very, very small piece of it because I know how to do it. You can still generate really great efficiencies, but it's just not the experience I think that everyone was looking for and desiring. But math still worked out. What you can do today is completely different. automate a much, much higher percentage of tasks, the long tail, the really complex narrows. But really beyond that, it's that you can orchestrate a very wide variety of scenarios for your agents to handle. And as a result of that, you can actually improve and really do champion challenger type tactics to really improve the outcomes of your portfolios, which just was not possible before because of, again, the level of training and level of specificity. You have to give your human agents to do the work. Now you can just like say, hey, 100 loans. I want to try these different strategies. I want to improve these different outcomes. I want to give these different notifications. Let's see what happens. Just completely different world. Yeah, which makes your job as a servicer a lot more fun. You've got some interesting things on the consumer side, too. you've been experimenting with? Yeah, so I would say the really fun part, thing that we get to do now and everyone gets to do now is you get to also just rethink how humans interact with systems, right? And so whereas before we were laser focused on building the right system of record and the way that humans interact with that system of record is through defined UIs. And then you actually end up hiring human operators whose skill set is knowing how to navigate the UIs. Now it's like we get to do things like we just build our own harness on top record and actually you just have, you know, that classic chat interface. And you get to do this on the back office side. And then when you think about how our homeowners interact with Valon OS, it's like you can do it through a voice AI, you can do it through a chat AI. There are just so many different form factors now. And it's really cool to think about what it could be. Awesome. And then maybe, Andrew, I think you, well, I know you Moonlight is one of the most productive on the roadmap or things you're thinking about solving. So there's the classic set of problems when it comes to, as you're building more and more of an enterprise, a really industry-specific harness, like decide, do you want to do fine-tuning? Do you want to improve your model efficiencies? Where do you want to invest in terms of a harness? So we've done the classic stuff when it comes to making sure you have vector databases, when you're trying to make sure you're pulling the right context, what is the sequence of context you want to put in. And then there's slight variations of this, which are very... industry specific. So in mortgages, as example, you care a lot more about correctness. And so you would probably want to align the models as much as possible and really the overall harness outcomes so that if you don't know the answer, it generally says, I don't know versus let me give you some hallucinated answer. Right. And so there's like small things here and there that really can change how the experience is for the user. But I think the The thing I find most fun and really the thing that I think is so amazing about the way that we are going about it as an industry or really as a product suite is marketing. Creating the infrastructure so that it's a continual learning process, right? So when you think about a lot of the infrastructure that people build and, you know, they're using, obviously, quite a bit of data from the user input and whatever else. That makes sense when you're doing coding. That makes sense when you have a pretty contained environment. But these are like really complex ecosystems where there's so many different components. And so these things change quite dynamically. And that's a very, very challenging problem because you can't say, "Well, here's the [30:00] And this is how we're going to solve it and we're going to optimize over a very, very clear function. Building a system that allows you to continuously run the evals, pull the data out, make sure you're actually continuously improving the model in a direction that is aligned, not just with what servicing or the way that servicing should be done generally, but customer specific. That is, I think, the part that's super interesting, building a overall set of infrastructure that can handle it at that capacity. Yep. Which segue as well, I know, even from the beginning, you both had ambitions broader than mortgage servicing. How does maybe the architecture and how you've set up the Valen platform set up Valen to expand into other sectors? Maybe I'll start off with a little known fact, which is servicing is like the nexus of all these different highly regulated enterprises. The fun fact I always like to go through with people is that people think servicing must be residential or maybe loan servicing. And you're like, okay, we talked about Resi and Resi is really great because it's regulated. It's the space that we started out in. It's very impactful because of the size of the market. And it's really near and dear to everybody's hearts, right? It's a home, American homeownership. It's really important. But you can always take one step left or right, whichever direction you want. And if it's commercial, there's all the data and all the information associated with the real estate, the tenants. Actually, people don't realize this, but as a commercial mortgage servicer, you actually get the information about the tenants' financials and all the data associated with that. So you could say, you know how like Toast and a bunch of people used their platforms to figure out how to do small business financing? It's like, you could kind of do the same thing if you decide to go through commercial. But if you take another step and you say, okay, let me think about this even more broadly, then you even have things like you know, healthcare, where there's revenue cycle management companies. People get very excited about those type of companies. And it turns out revenue cycle management is just servicing for hospitals and their claims and, you know, they're handling their patients. And the data that you get from that is, you know, effectively... all electronic medical records. And so it's a very exciting thing to do servicing first because it is a very, very, very challenging and difficult problem. It's very high breath. But when you get into it, you're in a position where you're effectively, you have tentacles across the entirety of the business and every element of what they work with and what they end up doing. And so it just gives you that really expansionary opportunity to say, hey, I want to go help you, you know, not just in making your servicing work, but I want to help you contact your customers. Let's just go back to Resi for a second. People, again, think of ResiServicing as, oh, I'm going to help you collect the payments and interact with the homeowners. Well, whatever we turned it into, I'm going to figure out a way to use this relationship that you guys have with all of your homeowners, and I'm going to call them every single month. I'm going to build a relationship with them. I'm going to make sure I have a pulse on what's going on with them every single month. And that's never been done before. And that's not how people think of servicing. But with a mixture of the technology that we've built and really where AI is headed, you can do those things. And I think that's what I find so, so exciting about this. Everything is servicing. Yes, everything is servicing. That could be the next step. Exactly. It kind of underpins so many things. Yes, it is the tentacle that like, or I don't know what you want to call it, but... It's the underlying critical infrastructure that supports basically anything that has money movements, some sort of regulation, and then an operational component. the stickiest, it's actually the hardest and most complicated as well. And so, you know, we thought, why not do that one first and then we'll go conquer the rest? No, it gets easy from here. Exactly. Perfect. So I want to tell a little bit about where Valen is today. So you built the servicer, recently announced, sold the servicer to sell the software platform. And then just recently, Rhythm confirmed that they're going to transfer four million loans over to Valen, which I think is the largest servicing transfer. Exactly. So and they, you know, they're smart. They baked you off across all the incumbents. What gave them a conviction to make such a big move? six months of us officially going to market as a software company, we signed over $200 million of deals. And that just tells you how much the industry has been wanting change. And so, from New Res' perspective, they've partnered with us really closely from the beginning. So, Valen Mortgage, subserviced New Res Loans, they could see the performance in a very clear, almost like demo room, basically, of this is exactly the performance that you get if you run software. Actually, I would argue that's the lower bound of the performance that you get if you run on the software. And then the way that we kind of structure all of our customer engagements [35:00] We want to make sure that we're incentive aligned with the customer. I think the second that you lose that alignment, that's when you're kind of going downhill as a company. And so for us, it's like we structure everything as outcomes oriented. And so it's for every dollar that our customer earns or for every dollar that our customer saves. That's when we earn as well. And I think that core alignment was the thing that made it just a no brainer for all of our customers. Maybe I want to end on, and anyone who's in New York, highly encourage them to visit the Valen office. This has got a ton of energy, but how would you describe what it's like to work at Valen, or what might people find surprising? Yeah, I think, look, Plenty of people have really great talent at their companies. There's really smart people obviously working at a variety of different companies. But I think what's really unique about the people at Valon is they're all the type of people where they want to take matters into their own hands. And I've had this conversation with so many of the people who've both joined and I've asked people who joined prior to us actually figuring out why people were joining. But, you know, these people, if you look at the people who've joined Valon, they have previously worked at or they had offers to go to work at, you know, obviously the standard thing companies, but OpenAI, Anthropic, all those type of things. And what makes a difference for them is the fact that they realize Look, there's a lot of smart people working on these other problems. Those will get solved. Clearly, like at the compute and the underlying model improvements, it's happening, it's working. Are you going to be the person who actually solves this problem, this really, really important problem? And finding the desire and finding joy in the fact that you're personally driving that impact in a way that you can so clearly and actually handle or deal with. That is, I think, what really, really gets people excited. And then... they realize they join for the people because they end up realizing, okay, besides the fact that I love doing this type of work and being very hardcore and making sure we're fixing these really, really challenging, impossible problems. You're surrounded by other people. Yeah, it's just like the same way. The people really make, people stay very long. Really, when we look at the way that our company has been built and formed, I mean, I want to say like their entire... Probably 75-80% of our management team is filled with people who've been here for five plus years. Yeah, exactly. It's really a demonstration of the type of company you have to build, the type of culture you have to build, and really the type of people who thrive and work at Valen. What would you say on the deployment side? I would say, and we're kind of going through this learning curve live, obviously. enterprise. It is... Unlike what you would imagine if you were just sitting there thinking about it in the abstract. Because at the end of the day, it's like what we're talking about is really large organizations of thousands and thousands of people. Every single individual has a slightly different objective function. You have layers of the organization that you need to move through. If you had asked me... six years ago, I would have told you, you know, this is a technology problem. What I know today is that this is a change management problem. And so what we've spent the last six to 12 months building is really that change management muscle of just how do you navigate organizations? How do you make change happen? And it turns out change is really, really hard at scale. But it is, I think, for this moment that we're in with AI, the number one biggest value driver for the next decade. And you've made the last point of that. You've already done a couple of successful deployments. Like what is the type of person that tends to really thrive in that kind of role? on what the objective function is, and then figure out what is the global thing that everyone's trying to do, bridge that, and then actually come up with a solution. It's really, I would say, usually there are people who are really, really great problem solvers. And then there are also people who are really, really good with dealing with people. And you need both. Yeah. Fantastic. Well, everything is servicing. Andrew, Linda, thank you for joining the podcast. Thanks for having us. Thanks, Angela. Thanks again for listening, and I'll see you in the next episode. [40:00] or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures. Thank you. Thank you. you you Thank you.