Panel 2 - 1: Jesse AI: Community-Centered Pilot Evaluation of an AI Chatbot for Substance Use

Panel 2 - 1: Jesse AI: Community-Centered Pilot Evaluation of an AI Chatbot for Substance Use — Ai4

Ai4

AI-assisted summary of Panel 2 - 1: Jesse AI: Community-Centered Pilot Evaluation of an AI Chatbot for Substance Use — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 19, 2026, this presentation features East Tennessee State University DrPH candidate and Backyard AI consultant Nicholas Swope. It explores the design, safety guardrails, and pilot evaluation of Jesse AI, a 24/7 anonymous chatbot created alongside Recovery Resources Tennessee to assist men in sober living with cravings, house rules, and community resources.

## TL;DR Nicholas Swope presents the implementation science behind Jesse AI, an anonymous RAG-powered chatbot developed for a men's sober living facility in Northeast Tennessee. Built to bridge 24/7 staffing gaps and relieve repetitive administrative burdens without replacing human peer connection, the platform incorporates local vernacular, safety off-ramping, and verified regional recovery resources. Evaluation data from a 7-day pilot involving 125 residents demonstrated strong engagement, favorable usability ratings on the Mobile Application Rating Scale, and particular utility for newly arriving residents. To ensure long-term sustainability and authentic adoption, the architecture features a no-code update pipeline for non-profit staff and a cultural voice filter modeled on staff interviews.

## Atmosphere & style The presentation carries a grounded, compassionate, and pragmatic tone that balances public health research with authentic grassroots recovery advocacy. Swope speaks with transparent humility about technical crashes and initial design missteps, grounding technical AI concepts in the lived experiences of residents. The delivery feels purposeful, earnest, and deeply respectful of the community, prioritizing human dignity and practical utility over flashy technological novelty.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=4s) Nicholas Swope introduces Jesse AI as an applied AI chatbot developed through implementation science to create sustainable, mission-aligned tools for resource-constrained substance recovery environments. - [01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) Recovery Resources Tennessee, a peer-run nonprofit serving 169 men in sober housing, faced critical operational bottlenecks including staff time constraints, lack of 24/7 staffing, and repetitive navigation queries from newly released residents. - [04:40](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=280s) A complex, personalized chatbot with user logins and EHR integration was rejected by staff and residents due to institutional distrust, leading the team to adopt a fully anonymous, conversational architecture. - [06:14](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=374s) Jesse AI integrates a curated knowledge base of house rules and 12-step literature, 398 verified regional resources in Northeast Tennessee, local vernacular tuning, and automated safety off-ramps for crisis situations. - [08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) During a 7-day pilot with 125 active residents out of 169 eligible men, Jesse AI processed roughly 1,500 queries across 700 conversational sessions, averaging two queries per session. - [08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) Residents used Jesse AI across varied recovery stages, from new residents navigating house rules and meeting locations to senior peer mentors using it as a sponsorship and craving management toolkit. - [10:50](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=650s) Evaluation via the Mobile Application Rating Scale (MARS) yielded an overall rating of 'good,' with users praising rapid response latency and around-the-clock availability despite high launch traffic causing an initial server crash. - [12:44](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=764s) Data triangulation revealed high active engagement and strong utility, with senior residents noting the bot would have served as an invaluable onboarding lifeline during their early transition into sober living. - [14:02](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=842s) To ensure long-term sustainability, Jesse AI utilizes a RAG architecture that allows non-technical nonprofit staff to update policies and local resource directories directly without coding expertise. - [15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) An initial deployment failure caused by generic, robotic phrasing was resolved by conducting qualitative interviews across all staff and founders to build a cultural voice filter into the response generation pipeline.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=4s) Introduction and Implementation Science in Recovery

Nicholas Swope introduces Jesse AI as an applied artificial intelligence tool designed specifically for a mission-driven substance use recovery organization. Drawing on more than a decade of experience as a public health practitioner, his current doctoral work in public health and AI applications at East Tennessee State University, and his role as lead consultant for Backyard AI, Swope frames the project around implementation science—which he colloquially defines as "the jargony term to say how do we do stuff that sticks?" Rather than focusing purely on technical novelty, the objective is understanding how to build, integrate, and sustain practical technology within non-tech community environments.

The presentation is structured into four core areas: introducing Jesse AI, sharing authentic stories from residents in recovery, presenting preliminary empirical results from the pilot implementation, and detailing an organizational adoption framework.

- Speaker background: Over 10 years as a public health practitioner; DrPH candidate at East Tennessee State University studying AI application; Lead Consultant at Backyard AI. - Presentation roadmap: Jesse AI overview, recovery case studies, pilot evaluation data, and a replicable implementation framework.

---

[01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) Recovery Resources Tennessee and the Pain Points of Care

The pilot was conducted in the epicenter of the Appalachian overdose crisis in the far northeast corner of Tennessee. Swope partnered with Recovery Resources Tennessee, a peer recovery support nonprofit where every staff member and resident is in personal recovery. The organization serves 169 resident men living in sober housing, almost all of whom were recently released from incarceration. Swope emphasizes that if a technology solution can succeed in this high-need, low-resource environment, it can be replicated anywhere.

About 13 months prior, Swope sat down with the organization's leadership not with an immediate technical pitch, but to ask fundamental diagnostic questions: what was working, what was failing, and what administrative burdens were "stealing your soul" and pulling staff away from their mission. The organization identified three primary bottlenecks:

```mermaid flowchart TD accTitle: Organizational Pain Points at Recovery Resources Tennessee accDescription: Flowchart illustrating the primary operational challenges leading to the need for Jesse AI. subgraph Bottlenecks["Core Organizational Challenges"] A["Staff Time Deficits<br/>Spreading small staff thin with repetitive inquiries"] B["24/7/365 Coverage Gaps<br/>Inability to staff sober houses around the clock"] C["High Resident Navigation Load<br/>Post-incarceration men asking recurring entry questions"] end Bottlenecks --> D["Risk to Human-Centered Culture<br/>Need tools that assist without replacing interpersonal touch"] ```

- **Staff Bandwidth Constraints**: Staff members were spread exceedingly thin answering basic administrative and navigation questions throughout the day. - **Coverage Deficits**: The nonprofit lacked the funding and capacity to provide physical staffing 24 hours a day, 7 days a week, 365 days a year. - **Repetitive Basic Queries**: Men newly transitioning out of incarceration into sober living had hundreds of questions; while critical issues required staff intervention, the vast majority were identical, recurring queries about house rules and local resources. - **Preserving the Human Touch**: Any proposed intervention had to preserve and enhance, rather than replace, the interpersonal connection central to peer recovery.

---

[04:40](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=280s) Chatbot Architectures: Why Complete Anonymity Won

When Swope initially proposed a sophisticated, deeply integrated chatbot 13 months prior, the staff reacted with immediate skepticism, hearing it as "too much, too fast" and fearing an "AI takeover." In a population with deep-seated institutional distrust toward both government and technology, a complex system would fail. Swope conceptualizes chatbot architectures across three general categories:

```mermaid flowchart LR accTitle: Spectrum of Chatbot Architectures accDescription: Three-tier spectrum comparing heavily personalized, semi-personalized, and fully anonymous chatbot architectures. subgraph Spectrum["Chatbot Integration Spectrum"] A["Heavily Personalized<br/>• EHR integration<br/>• Mandatory login<br/>• Warm handoffs"] B["Semi-Personalized<br/>• Optional accounts<br/>• Partial data linkage<br/>• Limited handoffs"] C["Fully Anonymous<br/>• Purely conversational<br/>• Zero name/identity tracking<br/>• Localized knowledge base"] end A -. Rejected: Created distrust.-> C B -. Bypassed for simplicity.-> C C ==> D["Selected Model: Jesse AI"] ```

While Swope originally envisioned a heavily personalized architecture with user logins and warm clinical handoffs similar to an electronic health record (EHR) assistant, community feedback proved that this approach would alienate residents. Instead, the team selected a completely anonymous, purely conversational model. By eliminating user tracking and authentication barriers, the tool respected resident privacy while delivering immediate utility.

---

[06:14](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=374s) System Design, Safety Off-Ramping, and Localized Knowledge

Jesse AI was developed through three successive rounds of staff co-design and pilot testing. The platform was built using a Retrieval-Augmented Generation (RAG) library hosted on an accessible platform rather than custom hard-coded software, ensuring nonprofit staff can independently update and maintain the system long after the academic research concludes.

```mermaid flowchart TD accTitle: Core System Components of Jesse AI accDescription: Diagram showing the four foundational architectural elements of Jesse AI. subgraph Engine["Jesse AI Architecture"] K["Curated Knowledge Base<br/>Policies, house rules, 12-step literature"] V["Regional Vernacular<br/>Trained on Northeast Tennessee dialect and recovery idioms"] R["398 Local Verified Resources<br/>Geo-located AA/NA meetings and basic needs services"] S["Safety Off-Ramps<br/>Automated triage for cravings, self-harm, and crisis"] end Engine --> Out["24/7/365 Anonymous Resident Support"] ```

The system is anchored by four key structural elements: - **Curated Knowledge Base**: Houses organizational policies, standard house rules, and literature for their 12-step recovery model. - **Localized Regional Vernacular**: Tailored to recognize the specific conversational patterns, slang, and dialect of Northeast Tennessee residents. - **398 Verified Regional Resources**: Hyper-localized community directory (e.g., looking up Alcoholics Anonymous meetings specifically in Johnson City, Tennessee, rather than generic national directories). - **Safety Off-Ramping**: Context-aware safety triggers that instantly identify statements involving cravings, suicidal ideation, or intentions of harm to self and others, directing users to emergency support.

---

[08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) Pilot Usage Data and Resident Stories: Justin and Danny

The quantitative pilot, conducted as part of Swope's doctoral dissertation, evaluated system usage over a 7-day monitoring window across the facility's population. Out of 169 eligible residents, 125 men voluntarily opted into the pilot. During that single week, the bot processed roughly 1,500 individual queries across approximately 700 distinct conversational sessions—averaging about two queries per interaction stream before the resident successfully resolved their inquiry.

Beyond aggregate analytics, Swope highlighted the lived experiences of two residents:

- **Justin**: A newer resident who utilized Jesse AI primarily for orientation and logistics, using hyperlinked GPS integrations to navigate to local 12-step meetings, as well as looking up internal sober living house rules and organizational policies. - **Danny**: A resident with two years of sustained recovery who serves as a house leader and mentor. Danny used Jesse AI not only to manage his own personal recovery cravings, but as an active sponsorship and mentorship toolkit to help junior housemates navigate acute cravings, find external social services, and organize 12-step sponsorship plans.

---

[10:50](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=650s) App Evaluation (MARS) and Qualitative Focus Group Findings

Following the week-long deployment, residents completed the Mobile Application Rating Scale (MARS), evaluating Jesse AI across four standardized dimensions: engagement, functionality, aesthetics, and information quality. On the MARS 1-to-5 scale (where 1 is completely inadequate and 5 is excellent), residents anonymously assigned an overall rating of "good."

In qualitative focus groups, Swope gathered targeted feedback on what functioned well, what failed, and what required refinement:

- **What Worked Well**: Residents praised the rapid response latency, instant access to accurate meeting locations, frictionless navigation to basic community needs, and around-the-clock availability. - **Challenges and Failures**: Residents identified specific resource coverage gaps where information was missing or outdated. Swope noted that high initial launch traffic actually caused a technical system crash, requiring an emergency mid-launch scramble to restore server availability.

[12:44](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=764s) Triangulating Pilot Data and Value for New Residents

Synthesizing user feedback, focus groups, and survey metrics revealed several critical insights into how Jesse AI functions in practice. Initial user reactions highlighted areas for improvement—specifically addressing the overly homogeneous tone across responses, expanding the underlying knowledge base, and refining the conversational user experience. When triangulating these preliminary data streams, three major conclusions emerged: residents are actively using the application, individual survey responses validate that the tool performs reliably, and it holds exceptional value for newcomers entering recovery housing.

Focus groups surfaced an enthusiastic undercurrent among senior residents who had already spent a year or two in the program. Reflecting on their own early days in recovery, these veterans commented, "Man, this is cool, but I would have loved this being new." This strong endorsement highlighted the chatbot's potential as an onboarding lifeline, easing the transition for incoming residents navigating unfamiliar rules and support systems during their vulnerable first weeks.

- **Key Preliminary Triangulation Insights:** - High resident engagement and active real-world usage. - Favorable individual survey evaluations confirming functionality. - Disproportionate value for newly arrived residents adjusting to the sober living environment.

---

[14:02](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=842s) Implementation Science and Non-Technical Knowledge Base Management

From an implementation science standpoint, long-term sustainability hinges on empowering the non-profit's internal team. Because policy manuals, house rules, and recovery resources evolve continuously, the technical architecture must allow non-technical staff to make seamless updates. A major design priority was building an accessible administrative pipeline where staff can update policy documents and RAG library resources directly without requiring specialized technical expertise.

Ensuring non-profit staff can maintain instant data synchronization ensures the chatbot's knowledge base reflects the latest organizational policies in real time. This operational integration bridges the gap between high-level technological capability and day-to-day grassroots utility, keeping the tool maintainable within resource-constrained recovery environments.

```mermaid flowchart TD accTitle: Jesse AI Knowledge Base and Voice Filter Architecture accDescr: Flow diagram showing how non-profit staff updates feed the RAG library, which passes through a cultural voice filter before delivering responses.

subgraph StaffAdmin["Staff & Knowledge Management"] A["Non-Profit Staff\n(Non-technical)"] -->|"Update rules & policies"| B["RAG Knowledge Base\n& Policy Library"] end

subgraph PromptEngine["Jesse AI Response Pipeline"] C["User Query"] --> D{"Voice & Persona Filter\n(Founders & 8 Staff Profiles)"} B --> D D --> E["Contextualized Recovery\nSupport Response"] end ```

---

[15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) Overcoming Early Failure Through Cultural Voice and Persona Design

The initial pilot of Jesse AI roughly nine months earlier struggled significantly upon release, essentially falling on its face with staff and residents. Qualitative feedback revealed why: the generic AI lacked the distinct voice, persona, idioms, and relational tone native to the organization. Without these shared cultural markers, the tool felt sterile and disconnected from the community it sought to support.

To remedy this disconnect, Swope conducted in-depth qualitative interviews with both co-founders and all eight staff members at Recovery Resources Tennessee. These narratives and behavioral patterns were distilled into a specialized voice filter. In the revised architecture, whenever Jesse retrieves factual information from the knowledge base, it passes the data through this cultural voice layer first. This ensures every output resonates with the warmth, recovery terminology, and authentic persona of the house staff before reaching the user.

- **Initial Launch Obstacle:** Failed user adoption due to robotic, generic chatbot phrasing that lacked organizational identity. - **Formative Data Collection:** Comprehensive qualitative interviews conducted across both founders and all 8 staff members. - **Architectural Adjustment:** Implemented a two-stage retrieval pipeline passing raw knowledge through an organizational voice filter prior to final output generation.

---

[16:34](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=994s) Acknowledgments and Collaborations

The presentation concluded with deep gratitude for the frontline workers and leadership at Recovery Resources Tennessee, whose dedication makes community-driven recovery work in northeastern Tennessee possible. Swope welcomed outreach from other mission-driven organizations, technologists, and academic researchers seeking to collaborate, replicate the implementation framework, or explore future studies in technology-supported recovery.

## Notable quotes > It's 2 a.m. in a sober living house; a resident is fighting a craving, and no staff is on site. Who answers? — Nicholas Swope

> [00:04](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=4s) the jargony term to say how do we do stuff that sticks? — Nicholas Swope

> [01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) stealing your soul — Nicholas Swope

> [04:40](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=280s) too much, too fast — Nicholas Swope

> [04:40](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=280s) AI takeover — Nicholas Swope

> [10:50](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=650s) good — Nicholas Swope

> [12:44](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=764s) Man, this is cool, but I would have loved this being new. — Resident

> keeps technology in service of human relationships, not in place of them — Nicholas Swope

## Mentioned Nicholas Swope, East Tennessee State University, Backyard AI, Recovery Resources Tennessee, Jesse AI, Justin, Danny, Alcoholics Anonymous, Mobile Application Rating Scale (MARS)

## Video description Nicholas Swope, Dr.PH Student, East Tennessee State University

It's 2 a.m. in a sober living house; a resident is fighting a craving, and no staff is on site. Who answers? In Northeast Tennessee's Appalachian region, where overdose deaths are high and recovery resources are scarce, I partnered with Recovery Resources Tennessee to design and pilot "Jesse AI," a chatbot that provides residents with 24/7/365, name-free support for cravings, house rules, 12-step questions, and job help. In this talk, I will share the community-centered design process, the safety guardrails that made it responsible, and fresh pilot data collected just weeks before the conference. If you want a replicable model for adopting AI in low-resource, mission-driven organizations that keeps technology in service of human relationships, not in place of them, this session is for you.

[Source](https://www.youtube.com/watch?v=aHzvRsjh2fs)

中文简报

专题讨论 2 - 1：Jesse AI：面向物质使用干预的社区中心型AI聊天机器人试点评估 Nicholas Swope 介绍了 Jesse AI 背后的实施科学。这是一款专为田纳西州东北部一家男性戒瘾康复居住机构开发的匿名 RAG 驱动聊天机器人，旨在填补全天候人员配备缺口并减轻重复的行政负担，同时不替代人际同伴支持。该平台融入了本土化语言风格、安全分流机制以及经核实的区域康复资源。一项涵盖 125 名居民的 7 天试点评估数据显示出强烈的用户参与度，在移动应用评分量表（Mobile Application Rating Scale）上获得了良好的可用性评分，对新入住居民尤为实用。为确保长期可持续性与真实采用，该架构还为非营利机构员工提供了无代码更新流程，并配备了基于员工访谈建模的文化语调过滤器。 Nicholas Swope 介绍了通过实施科学开发的戒瘾恢复聊天机器人 Jesse AI，旨在为资源匮乏的物质成瘾康复环境提供可持续、符合使命的技术支持。 (0:04) 面对机构信任缺失的问题，团队放弃了需要用户登录和电子病历（EHR）集成的复杂方案，转而采用完全匿名的对话式架构，以解决员工时间不足和无法提供 24/7 全天候支持等运营瓶颈。 (4:40) Jesse AI 整合了宿舍规章、十二步疗法文献、398 项经核实的田纳西州东北部本地资源、当地方言调优以及针对危机情况的自动化安全转介机制。 (6:14) 在为期 7 天的试点中，125 名戒瘾宿舍男性居民共发起约 700 次会话、处理了 1500 条查询；移动应用评定量表（MARS）总体评价为“良好”，尤其为新入住居民提供了关键过渡支持。 (8:52) 针对早期因语气机械而遇挫的教训，团队通过深度访谈将机构员工的语言风格提炼为文化声音过滤器，并搭建了便于非技术人员直接维护的 RAG 知识库更新流程。 (15:03)

Reading summary

Public health practitioner Nicholas Swope evaluated Jesse AI, a fully anonymous conversational chatbot tailored for a peer-run substance use recovery organization in Appalachian northeast Tennessee. The pilot demonstrated that integrating regional vernacular and empowering non-technical staff to manage the knowledge base enabled an artificial intelligence tool to resolve round-the-clock administrative bottlenecks without eroding peer trust. Operational Strains in Sober Living Facilities (1:48) Nicholas Swope, a public health practitioner and DrPH candidate at East Tennessee State University, partnered with Recovery Resources Tennessee, a peer-led nonprofit where all staff and residents are in personal recovery. The organization operates sober housing for 169 men, nearly all recently released from incarceration. Facing high resident intake in a low-resource environment, leadership reported acute operational constraints: a small staff was constantly consumed by repetitive daily logistical questions, and the facility lacked the budget and capacity to provide physical staffing 24 hours a day, 7 days a week, 365 days a year. Incoming residents faced hundreds of recurring questions regarding house rules, local recovery resources, and daily navigation. While urgent clinical issues required staff attention, the continuous stream of mundane queries pulled workers away from direct mentorship. Swope noted that any viable technological intervention needed to alleviate this administrative burden while safeguarding the interpersonal peer connections central to the community's recovery culture. Designing for Trust: Anonymity and Cultural Voice (4:40) Initial proposals for an electronic health record (EHR)-style system with user accounts encountered immediate pushback from staff and residents, who feared surveillance and technological intrusion. Recognizing widespread institutional distrust, Swope abandoned personalized tracking in favor of a completely anonymous conversational model built on Retrieval-Augmented Generation (RAG). The system incorporated automated safety off-ramps to triage acute cravings, suicidal ideation, or self-harm toward emergency support, paired with a localized database of 398 verified regional resources, such as specific 12-step meetings. An earlier version of the bot deployed nine months prior had failed because its sterile, robotic phrasing lacked the organization's identity. To rectify this, Swope conducted qualitative interviews with both co-founders and all eight staff members. The team translated these insights into an organizational voice filter, routing all factual knowledge base outputs through the distinct regional dialect, idioms, and relational warmth characteristic of northeast Tennessee recovery communities before delivering responses. Pilot Evaluation Metrics and Dual-Use Patterns (8:52) During a seven-day quantitative dissertation pilot, 125 of the 169 eligible residents voluntarily used Jesse AI. The system logged approximately 1,500 queries across roughly 700 conversational sessions, averaging about two queries per interaction before resolving the user's issue. On the Mobile Application Rating Scale (MARS), residents anonymously evaluated the tool across engagement, functionality, aesthetics, and information quality, yielding an aggregate score of "good." Focus groups revealed that while senior residents praised the application, they noted it offered the greatest onboarding utility for newly arrived residents adjusting to house rules. Individual case studies illustrated differing usage patterns across recovery phases. Justin, a newer resident, used GPS integrations to locate external 12-step meetings and review internal house regulations. Conversely, Danny, a mentor with two years of sobriety, utilized the chatbot both to manage personal cravings and as a sponsorship toolkit to locate social services for junior residents. Reported drawbacks included an initial server crash caused by launch-traffic spikes and isolated coverage gaps in the resource database. Implementation Science and Non-Technical Maintenance (14:02) Swope framed the project around implementation science, emphasizing how to integrate practical technology so it remains functional in non-technical settings over the long term. Because community resources, organizational policies, and house manuals change continuously, the technological framework was designed to prevent reliance on outside software developers. Rather than embedding rules in hard-coded software, the platform utilizes an accessible administrative pipeline that allows non-profit staff to update RAG documentation directly. Swope argued that ensuring non-technical frontline personnel can independently modify policies and synchronize data in real time is vital for sustaining community-level artificial intelligence tools once academic research concludes. The Jesse AI pilot demonstrates that an anonymous, culturally adapted RAG chatbot can mitigate staffing deficits in low-resource recovery housing, particularly for onboarding newly released residents. However, Swope noted lingering operational challenges, including preventing server instability during traffic spikes, diversifying homogeneous response phrasing, and continuously maintaining local resource directories.

精炼阅读摘要

公共卫生实践者兼东田纳西州立大学博士候选人 Nicholas Swope 介绍了 Jesse AI 的试点评估结果。该项目将基于检索增强生成（RAG）的匿名对话工具引入阿巴拉契亚地区的戒瘾康复机构，以填补资源匮乏环境下的服务缺口并减轻员工行政负担。 高需求康复机构的服务痛点与技术诉求 (1:48) 该试点在田纳西州东北部阿巴拉契亚药物过量危机重灾区展开，合作机构为同行康复支持非营利组织 Recovery Resources Tennessee。该机构的所有工作人员和居住者均处于个人康复阶段，服务于居住在戒瘾公寓内的 169 名男性，且几乎所有人员近期都有出狱经历。Swope 指出，若技术方案能在这种高需求、低资源的极端环境中成功落地，便具备在其他场景复制的可行性。 在项目启动前的诊断中，机构管理层提出了三项核心运营瓶颈：一是工作人员有限的精力被全天大量重复的基础行政和导航咨询严重分散；二是受资金与人员编制所限，无法实现全天候 24/7/365 的物理驻留覆盖；三是新入住人员因重新适应社会而产生海量且重复的规则与本地资源疑问。技术方案的核心约束在于，必须在缓解事务性负荷的同时保留并强化同行康复至关重要的人际联结，而非用技术取代人工关怀。 架构权衡：完全匿名性与本地化知识库 (4:40) 面对服务群体对政府和技术机构根深蒂固的不信任感，团队最初构想的深度整合方案遭到抵制。Swope 对比了三类架构：一是需要强制登录并打通电子病历（EHR）的高定制型，因引发对数据追踪的恐慌和“AI接管”担忧而被否决；二是可选账户的半个性化型；三是完全不追踪身份、仅提供本地知识库的纯对话匿名架构。最终选用的完全匿名模式彻底移除了身份验证障碍，从而在保护隐私的前提下迅速建立了信任。 Jesse AI 的底层系统依托可访问平台的检索增强生成（RAG）库构建，确保非技术人员在研究项目结束后仍能自主维护。系统核心包含四大支柱：容纳机构政策、房屋管理规则及十二步骤戒瘾文献的精选知识库；识别东北田纳西方言与康复特定习语的本地化语料库；涵盖 398 个经过验证的本地资源名录（如精准到约翰逊城的 AA/NA 互助会地点与基本生活服务）；以及针对渴求、自残意图或危机言论的自动化安全分流机制，以即时转接紧急干预支持。 克服初期失败：融入组织文化与同行声音 (15:03) Jesse AI 在初次发布时遭遇重大挫折，未能获得员工和居民的接纳。质性反馈显示，早期的通用型 AI 输出语调冰冷刻板，缺乏机构内部特有的语言风格、关系温度与康复术语，从而产生了脱节感。Swope 指出，纯粹的技术有效性无法替代真实的文化共鸣。 为纠正这一缺陷，Swope 对机构的 2 位联合创始人以及全部 8 名员工进行了深度质性访谈，提炼出组织特有的叙事结构和行为表达模式，并构建了专属的文化语音过滤器。在调整后的双阶段检索管道中，系统从知识库检索出事实信息后，必须先经过这一组织语音层过滤，再生成带有员工温度与同行认同的回复，从而有效化解了早期因语气机械而引发的抵触情绪。 为期一周的试点量化表现与质性评估 (8:52) 在为期 7 天的博士论文试点监测期内，169 名符合条件的居民中有 125 人自愿加入。系统在约 700 次独立会话中处理了约 1500 条查询，平均每个会话仅需 2 次提问即可解决问题。移动应用评定量表（MARS）在参与度、功能性、美学和信息质量四个维度上的整体评分匿名归类为“良好”。个案反馈显示，新人 Justin 主要利用 GPS 超链接导航参加互助会并查阅管理规则；而已有两年康复经历的资深居民兼宿舍骨干 Danny，则将该工具用作指导新人缓解急性渴求及协助对接社会服务的重要辅助支持。 质性焦点小组研究同时揭示了优势与技术局限。居民称赞该工具响应迅速、全天候可用且本地资源定位准确，老住户更特别指出其对刚入住房客的巨大缓冲价值。但在初期的高并发访问中，服务器曾出现崩溃并触发紧急修复；此外，用户也指出了部分资源信息陈旧或缺失的问题，并对系统在多轮对话中过于单一的语调提出了改进意见。 面向可持续维护的实施科学管理管线 (14:02) 从实施科学的角度来看，基层非营利组织引入技术的关键不仅在于开发，更在于能否长期落地（“how do we do stuff that sticks”）。鉴于康复机构的管理制度、房屋规则和外部转介资源处于持续变动中，Jesse AI 特别设计了低门槛的管理操作管线，使完全不具备技术背景的机构员工能够独立、无缝地更新政策文档和 RAG 库。 这种知识库与运营实践的实时同步机制，使草根机构摆脱了对外部科研团队或专业程序员的技术依赖，填补了高端 AI 技术与严苛现实资源之间的断层。Swope 强调，只有将数据所有权与系统维护权切实交还给非营利组织员工，技术介入才能在资源匮乏的社区环境中实现真正的长期可持续运转。 试点表明，结合完全匿名、本地化知识检索和文化适配语音层的聊天机器人，能有效缓解戒瘾过渡安置机构的工作负荷。但该系统的长期成效仍受限于服务器并发承载能力、本地资源数据的持续更新频次，以及在扩展至其他社群时对独特方言与文化层的高额前期适配成本。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 19 日发布在 YouTube 平台。内容由东田纳西州立大学公共卫生博士候选人 Nicholas Swope 主讲，分享了他在阿巴拉契亚药物过量危机重灾区与非营利机构 Recovery Resources Tennessee 合作开发并试点“Jesse AI”聊天机器人的全过程，旨在为资源匮乏的使命驱动型组织探索负责任的 AI 落地模式。 ## TL;DR 面对阿巴拉契亚地区严峻的成瘾危机及康复机构夜间人手匮乏、学员对技术缺乏信任的痛点，团队采用实施科学方法打造了完全匿名的 24/7 对话助手 Jesse AI。在 7 天的集中试点中，125 名刚出狱的清醒宿舍住户发起了约 1500 次提问，数据与定性访谈证明其对初入康复期的新人尤其具有支持价值。该系统通过免代码的 RAG 知识库让非技术员工能够自主维护，并结合员工访谈提炼出独特的“声音过滤器”，成功化解了通用 AI 语气生硬冷漠的问题。 ## Atmosphere & style 演讲兼具学术研究的严谨与一线公共卫生实践的温情，Swope 的语言通俗生动且极其接地气。他毫不避讳早期原型上线彻底惨败的技术挫折，真诚复盘了如何一步步打破学员对技术的恐惧并重塑带有社区人情味的人设。整场分享洋溢着对底层康复群体与一线工作人员的敬意，传递出一种“让技术服务于人际温度而非取代人”的坚定使命感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=4s) 项目立足于实施科学（Implementation Science）框架，核心探索如何在资源匮乏的非技术、使命驱动型机构中成功落地并长期维系 AI 技术。 - [01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) 阿巴拉契亚地区的同行康复机构面临员工精力被琐事挤压、缺乏 24/7 全天候支持能力，以及亟需在解答海量重复问题时保留人际关怀温度的三大痛点。 - [04:15](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=255s) 针对戒瘾群体对科技和监控的天然戒备，项目放弃了重度个性化医疗方案，确立了无需登录、不记录个人身份的完全匿名对话架构以建立信任。 - [06:14](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=374s) Jesse AI 建立了严格的安全分流护栏（Safety Off-Ramping），能精准识别渴求与危机情况并触发干预，同时深度适配了当地语言与方言习惯。 - [06:14](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=374s) 系统基于 RAG 架构收录了 398 个经人工核实的本地戒酒/戒毒互助聚会资源，并提供一键式 GPS 导航链接与清醒宿舍规章查询功能。 - [08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) 7 天试点显示出极高的自愿采纳度（169 人中 125 人使用），平均每场对话仅需约 2 个轮次即可精准解决问题，并在 MARS 评估中获得良好评分。 - [08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) 学员不仅用 AI 应对偶发渴求和查询制度，资深康复者还将其作为“资助人计划”的辅助工具，辅导新学员践行“渡人渡己”的十二步法精神。 - [12:44](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=764s) 早期洞察显示 Jesse AI 对刚进入康复环境的新学员帮助尤为巨大，有效缓解了面对海量陌生规则时的焦虑感与适应门槛。 - [14:02](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=842s) 为确保系统长期可持续运行，管理后台被设计为极简的可视化界面，方便非技术背景的工作人员自主动态更新规章制度与社区资源知识库。 - [15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) 早期原型曾因通用 AI 语气生硬冷漠而失败，表明在心理防线脆弱的戒瘾群体中，缺乏同理心与人情味的算法无法建立信任纽带。 - [15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) 团队通过深度访谈全员提炼出独特的“声音过滤器（Voice Filter）”，构建起结合定性人设基调与 RAG 事实检索的双层生成架构，成功注入组织灵魂。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=4s) 介绍与背景：将 AI 引入非技术使命型组织 东田纳西州立大学（East Tennessee State University）公共卫生与 AI 应用方向的博士候选人 Nicholas Swope 带来了关于“Jesse AI”的实践分享。作为拥有十余年公共卫生从业经验的学者及 Backyard AI 的首席顾问，Swope 专注于实施科学（Implementation Science）与组织战略——用他的大白话来说，实施科学就是研究“如何让做的事情真正落地生根并持续发挥作用（how do we do stuff that sticks）”。本次演讲的核心主线围绕四个部分展开：Jesse AI 项目本身、成瘾康复中的真实故事、前沿试点数据以及一套可供非技术型使命组织复用的成功实施框架。 Swope 指出，他的目标不仅是展示一个已经在物质滥用康复领域落地的实用工具，更重要的是从学术和实践角度探讨如何在资源有限的“非技术领域”构建并应用技术，同时为其他非营利机构提供转型与落地洞见。 ### [01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) 阿巴拉契亚地区的危机与非营利机构的痛点 在展示美国疾病控制与预防中心（CDC）的药物过量致死地图时，Swope 指向了被称为危机震中（epicenter）的阿巴拉契亚山脉区域。在田纳西州东北角的深处，坐落着一家名为 Recovery Resources Tennessee 的同行康复支持非营利机构。该机构具有高度特殊性：从基层员工到管理人员全都是康复者本人，当前接纳了 169 名男性学员，其中绝大多数刚从监狱释放。Swope 直言：“如果你能在这里成功实施一项技术解决方案，你就能在任何地方做到。” 大约 13 个月前，Swope 第一次坐下来与该机构沟通。当时他并没有预设要做一个聊天机器人，而是直截了当地询问机构痛点：“目前什么有效？什么无效？最重要的是，什么事情正在消耗你们的精力、夺走你们的热情（stealing your soul）？”机构团队总结了三大核心挑战： 1. **员工时间被极度挤压**：作为小型非营利机构，人手捉襟见肘，大量时间耗费在回答日常琐碎问题和基础资源导航上。 2. **缺乏全天候支持能力**：机构无法做到“全天候 24/7/365”有人值班，但学员的成瘾渴求（cravings）和危机往往发生在深夜或无人在场时。 3. **重复性问题泛滥但需保留人情味**：出狱后住在清醒宿舍（sober living）的学员急于开启人生新篇章，面临海量疑问。部分重大问题确实必须由员工介入，但绝大多数都是千篇一律的基础咨询。更关键的是，他们绝不希望任何技术方案取代该机构赖以成功的“人际关怀与温度（human touch）”。 ### [04:15](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=255s) 从“技术接管”恐惧到完全匿名的架构设计 当 Swope 首次提出可以用聊天机器人建立一套完善的基础设施时，机构的第一反应是强烈的抗拒和恐惧，直呼“太快太过了（too much, too fast）”。学员和员工普遍对科技、政府以及外界抱有深层的不信任感，认为这是“AI 接管一切”。 为了打消疑虑并找到契合点，Swope 对聊天机器人的架构进行了分类梳理： - **重度个性化架构（Heavily Personalized）**：常见于医疗健康系统（如电子病历集成），需要学员登录并支持深度的热线转接（warm handoffs）。虽然这是 Swope 最初设想的方案，但在该场景下完全行不通，因为强制登录会直接劝退学员。 - **半个性化架构（Semi-Personalized）**：允许用户自主选择是否登录，提供有限的个性化服务与转接。 - **完全匿名对话架构（Completely Anonymous Conversational）**：无需任何登录和身份追踪，纯粹基于自然语言交互。 最终团队确立了完全匿名的架构路线，并借助实施科学方法进行本地化定制，使技术真正服务于人，而不是给人带来压迫感。 ```mermaid flowchart TD accTitle: Jesse AI Chatbot Architecture Spectrum accDescr: Comparison of heavily personalized, semi-personalized, and completely anonymous chatbot architectures, highlighting the selected approach. subgraph Spectrum["聊天机器人架构分类与选型"] A["重度个性化架构\n(医疗病历/账号登录/深度转接)"] B["半个性化架构\n(可选登录/部分个性化)"] C["完全匿名对话架构\n(Jesse AI 最终采纳方案)"] end A -.->|"信任门槛过高\n容易引发抵触"| Fail["放弃"] C -->|"无痕使用/24小时响应\n贴合康复信任需求"| Success["成功落地并持续运行"] ``` ### [06:14](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=374s) Jesse AI 的核心架构、本地化资源与安全护栏 经过三轮与机构员工及学员的共创开发和试点测试，Jesse AI 最终形成了四个核心模块并全面上线： 1. **定制知识库（Knowledge Base）**：精准收录康复机构所需的全部规章制度与专业指导。 2. **人格化声音与语气（Persona/Voice）**：采用富有同理心且契合康复群体的口吻进行交互。 3. **安全分流护栏（Safety Off-Ramping）**：具备精准识别危机情境的能力，一旦监测到用户表达成瘾渴求（cravings）或自残、伤人意图，立即触发安全干预与转介机制。 4. **地方俚语与本土语言适配**：深度适配田纳西州东北部阿巴拉契亚地区的真实语言习惯和表达方式。 Jesse AI 的底层基于检索增强生成（RAG）库构建，并部署在可视化、免代码维护的平台上。这样设计的目的是避免技术依赖——即便 Swope 未来退出项目，机构员工也能自行更新和维护系统。 - **全天候匿名服务**：不记录任何个人姓名与隐私，提供 24/7/365 即时支持。 - **十二步法（12-Step）康复体系支持**：深度整合机构所采用的十二步戒瘾理念与问答指南。 - **398 个经过人工核实的本地资源**：收录了涵盖约翰逊城（Johnson City）等当地的匿名戒酒会（AA）聚会点，并附带可直接点击导航的 GPS 超链接。 - **宿舍守则与规章导航**：内置清醒宿舍的所有规约条例，随时解答学员日常疑问。 ### [08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) 初步试点数据、真实学员故事与定性反馈 这项研究是 Swope 博士毕业论文的核心组成部分。在为期 7 天的集中试点周期中，全院 169 名居民中有 125 人自愿参与了测试。系统累计产生了约 1,500 次提问查询，分布在约 700 场独立对话中，平均每场对话包含约 2 个轮次的问题，表明学员通常在两次交互内即可精准获取所需信息。 除了冷冰冰的数字，Swope 分享了两位具有代表性的学员案例： - **Justin**：刚进入项目的学员，主要使用 Jesse AI 查找十二步康复聚会。他通过系统内嵌的 GPS 超链接一键导航到聚会地点，并频繁查询清醒宿舍的日常规章制度。 - **Danny**：已在机构成功康复 2 年的资深学员。他不仅用 Jesse AI 管理自己的偶发渴求，更将其作为“资助人计划（Sponsorship Planning）”与日常辅导的工具，帮助同宿舍的新学员寻找资源、应对渴求，践行了十二步法中“渡人即渡己”的精神。 在量化评估方面，研究采用了移动应用评分量表（MARS，从参与度、功能性、美观度和信息质量四个维度打分，1分为极差，5分为极佳）。学员在完全匿名、独立的情况下对 Jesse AI 的整体质量给出了“良好（Good）”的评价。 而在定性焦点小组访谈（Focus Groups）中，Swope 收集到了最真实的反馈： - **受好评的亮点**：极快的响应速度、精准的聚会地点导航、即时满足基础生活需求。 - **暴露的问题与不足**：部分偏远资源存在覆盖盲区，部分数据需要动态更新；此外，系统刚上线时由于瞬间并发访问量过大导致服务器崩溃，迫使 Swope 在发布中途紧急抢修恢复。 ### [12:44](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=764s) 试点初步反馈与三项核心洞察：赋能康复“新人” 在对早期试点数据进行三角互证（Triangulation）分析时，团队梳理出了几项关键的用户体验反馈。最初的反馈显示，AI 的部分回复显得有些单一和千篇一律（Homogeneous），这促使团队进一步优化聊天机器人的语言风格与语气（Voice），并着手扩充其检索所依赖的知识库。综合焦点小组访谈与定量调查，演讲者提炼出了三项当前最有说服力的初步核心洞察：首先，清醒生活之家（Sober Living House）的住户确实在真实、高频地使用这款应用；其次，个人问卷调查表明住户普遍认为系统运行良好、切实有用；最重要的是，这款工具对刚进入康复项目的新人展现出了非凡的价值。 焦点小组访谈中出现了一条非常清晰的情感暗线：已经在该康复机构生活了一两年的“老学员”们纷纷感叹：“天哪，这太酷了！要是我刚来那会儿能有这个该多好。非常期待看到它在未来几个月的发展。”对于刚脱离成瘾环境、初入清醒生活之家的住户而言，面对海量陌生的规章制度和作息要求往往会感到不知所措，而 Jesse AI 恰好成为了他们随时随地消除困惑、缓解心理压力的数字助手。 - **三大初步核心洞察**： - 住户在日常生活中形成了真实的自主使用习惯。 - 个人评估问卷显示系统解决实际问题的表现十分优异。 - 对刚入驻康复机构的新住户（New Guy）助益尤为显著，有效降低了新环境适应期的信息与心理门槛。 --- ### [14:02](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=842s) 实施科学视角：非技术人员的知识库维护与动态更新 从实施科学（Implementation Science）的框架来看，技术的成功落地不仅取决于算法本身，更依赖于机构内非技术人员能否无缝维护底层系统。非营利组织的政策手册（Policy Book）、社区资源清单以及规章细则经常需要动态调整。为了让 Jesse 能够即时获取最新信息，团队将重点放在了降低系统管理门槛上。 演讲者强调，非营利组织的专职工作人员并非技术极客（Not tech people），因此系统架构必须设计得极易操作。管理端被打造成直观易懂的 RAG（检索增强生成）知识库管理界面，工作人员无需编写代码，便能随时上传和更新内部制度文件。这种设计确保了当机构发布新的管理政策或当地支持资源发生变动时，后台知识库能够迅速刷新，使 Jesse 能够在第一时间向住户提供准确、合规的答复。 --- ### [15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) 失败反思与人格化重塑：注入组织灵魂的“声音过滤器”与结语 演讲者回顾了项目历程中的一次重大挫折：大约在 9 个月前，Jesse 的早期原型版本在工作人员内部上线测试时“彻底搞砸了”（Kind of fell on its face）。深入的用户调研揭示了失败的核心原因——当时的 AI 缺乏人情味与组织特有的“声音与人设”（Voice and Persona），说话腔调完全没有体现出当地康复社区的文化特色、口头用语与关怀特质。冷冰冰的通用 AI 语气无法在脆弱的戒断群体中建立起信任。 为了彻底扭转这一局面，演讲者对 Recovery Resources 的 2 位联合创始人以及全部 8 名专职员工进行了深度访谈，系统性地提炼并数字化了机构的沟通风格、价值理念与常用话术。团队将这些定性的人格特征构建成一层独特的“声音过滤器”（Voice Filter）。当 Jesse 生成回复时，它会先通过这层声音过滤器规范语气基调，再结合检索到的事实知识库输出回答，从而兼具专业准确性与社区特有的温度。 ```mermaid flowchart TD accTitle: Jesse AI 双层生成架构 accDescr: 展示用户提问后，系统如何通过声音过滤器结合 RAG 知识库生成具有机构人设的回复。 subgraph Input ["用户交互"] Q["住户输入问题<br/>（戒断渴望 / 机构制度 / 12步法）"] end subgraph Processing ["双层处理机制"] VF["声音与人设过滤器<br/>（创始人与8名员工访谈提炼）"] KB["RAG 知识库 / 政策手册<br/>（非技术员工动态维护更新）"] end subgraph Output ["生成输出"] Ans["温暖、契合社区口吻且准确的回复"] end Q --> VF Q --> KB VF --> Ans KB --> Ans ``` 演讲在对合作伙伴的真诚感谢中进入尾声。演讲者特别向 Recovery Resources Tennessee 的全体团队致谢，盛赞他们在田纳西州东北部阿巴拉契亚地区开展的卓越救助工作，并表示“没有他们，这一切都不可能实现”。最后，演讲者向在场的学术界学者与使命驱动型非营利机构发出合作邀请，期待在实施科学研究与公益 AI 落地层面开展更多跨界合作。 - **人设重塑关键数据与举措**： - **访谈覆盖**：深度访谈了机构的 2 位创始人与全部 8 位员工。 - **双层架构逻辑**：采用“声音过滤器（定性人设）+ RAG 事实知识库”的协同生成机制，解决通用大模型语气生硬冷漠的缺陷。 - **致谢与展望**：向田纳西州东北部一线禁毒康复团队致敬，呼吁更多学者和非营利组织共同探索低资源环境下的负责任 AI 实践。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=4s) 如何让做的事情真正落地生根并持续发挥作用（how do we do stuff that sticks） — Nicholas Swope > [01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) 如果你能在这里成功实施一项技术解决方案，你就能在任何地方做到。 — Nicholas Swope > [01:48](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=108s) 目前什么有效？什么无效？最重要的是，什么事情正在消耗你们的精力、夺走你们的热情（stealing your soul）？ — Nicholas Swope > [04:15](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=255s) 太快太过了（too much, too fast） — Recovery Resources 机构团队 > [08:52](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=532s) 渡人即渡己 — Danny > [12:44](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=764s) 天哪，这太酷了！要是我刚来那会儿能有这个该多好。非常期待看到它在未来几个月的发展。 — 康复机构老学员 > [15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) 彻底搞砸了（Kind of fell on its face） — Nicholas Swope > [15:03](https://www.youtube.com/watch?v=aHzvRsjh2fs&t=903s) 没有他们，这一切都不可能实现 — Nicholas Swope ## Mentioned Nicholas Swope, 东田纳西州立大学 (East Tennessee State University), Backyard AI, Recovery Resources Tennessee, Jesse AI, 美国疾病控制与预防中心 (CDC), 阿巴拉契亚地区 (Appalachia), 约翰逊城 (Johnson City), 匿名戒酒会 (AA), 十二步法 (12-Step Program), 移动应用评分量表 (MARS), RAG (检索增强生成), Justin, Danny ## Video description Nicholas Swope, Dr.PH Student, East Tennessee State University It's 2 a.m. in a sober living house; a resident is fighting a craving, and no staff is on site. Who answers? In Northeast Tennessee's Appalachian region, where overdose deaths are high and recovery resources are scarce, I partnered with Recovery Resources Tennessee to design and pilot "Jesse AI," a chatbot that provides residents with 24/7/365, name-free support for cravings, house rules, 12-step questions, and job help. In this talk, I will share the community-centered design process, the safety guardrails that made it responsible, and fresh pilot data collected just weeks before the conference. If you want a replicable model for adopting AI in low-resource, mission-driven organizations that keeps technology in service of human relationships, not in place of them, this session is for you. [Source](https://www.youtube.com/watch?v=aHzvRsjh2fs)

Source quotes

[0:04] Nicholas Swope: 如何让做的事情真正落地生根并持续发挥作用（how do we do stuff that sticks） [1:48] Nicholas Swope: 如果你能在这里成功实施一项技术解决方案，你就能在任何地方做到。 [1:48] Nicholas Swope: 目前什么有效？什么无效？最重要的是，什么事情正在消耗你们的精力、夺走你们的热情（stealing your soul）？ [4:15] Recovery Resources 机构团队: 太快太过了（too much, too fast） [8:52] Danny: 渡人即渡己 [12:44] 康复机构老学员: 天哪，这太酷了！要是我刚来那会儿能有这个该多好。非常期待看到它在未来几个月的发展。 [15:03] Nicholas Swope: 彻底搞砸了（Kind of fell on its face） [15:03] Nicholas Swope: 没有他们，这一切都不可能实现

Source quotes

Nicholas Swope: It's 2 a.m. in a sober living house; a resident is fighting a craving, and no staff is on site. Who answers? [0:04] Nicholas Swope: the jargony term to say how do we do stuff that sticks? [1:48] Nicholas Swope: stealing your soul [4:40] Nicholas Swope: too much, too fast [4:40] Nicholas Swope: AI takeover [10:50] Nicholas Swope: good [12:44] Resident: Man, this is cool, but I would have loved this being new. Nicholas Swope: keeps technology in service of human relationships, not in place of them

Original transcript

[0:04] So today uh in the short bit of time that we have  um we're going to be talking a little bit about [0:10] Jesse AI which is a tech um uh solution and it's  an addition for a missiondriven organization. But [0:19] what I know by the end of this is that at first  you're going to be able to use Jesse AI. It's an [0:25] it's a a practical tool that's already being used  within our substance use recovery. Um hopefully [0:33] the more important things for me as an academic is  looking at the implementation science, looking at [0:38] how to build tech in a non- tech space. And third,  hopefully you'll leave for insights for your own [0:43] organization. I know all this um because I've been  a public health practitioner for over a decade. [0:50] Um, I'm finishing up my doctorate degree in public  health and AI application at East Tennessee State [0:57] University. Um, I spend a lot of time looking  at strategy and implementation for how that [1:04] connects into missiondriven organizations. Um,  I spend a lot of time in implementation science, [1:10] which is really kind of what we're doing here.  It's the jargony term to say how do we do stuff [1:15] that sticks? Um, I spend a lot of time looking  at different chat bots, how to build them, [1:20] their archetype, their architecture, and how  to actually bring that into organizations. Um, [1:27] and I'm the lead consultant for Backyard AI. So,  really quickly, if you forget things, we're just [1:32] going to do four things. The first one is we're  going to talk about uh Jesse AI. The second thing, [1:37] and the most important thing is we're talking  about stories in recovery. Third, we're going to [1:43] talk about some preliminary results from how this  has been implemented and then a success framework [1:48] that you should be able to have. A little bit  of context. Um, this is a CDC map of substance [1:55] use disorder or uh, excuse me, substance use  deaths in the US. That pink triangle is kind of [2:02] the Appalachian Mountains. I did a terrible job at  highlighting it. Um, but it's the epicenter in the [2:08] very northeast corner of northeastern Tennessee  is where I am and where this amazing organization [2:15] works. It's called Recovery Resources Tennessee.  It is a peer recovery support nonprofit. Everybody [2:22] from the staff to the individuals um are  all in recovery. They have 169 current men, [2:30] almost all of them recently out of incarceration.  If you can implement a tech solution here, [2:37] you can do it anywhere. So I initially sat  down about 13 months ago with the organization, [2:44] not thinking that I was going to do a chatbot.  We're sitting down and we're saying, okay, [2:49] what's working? What isn't working? And most  importantly, what's taking you away from your [2:55] missiondriven work? What's kind of stealing your  soul? And they said three kind of main things. [3:03] The first one which runs into almost all small  nonprofits which is staff time. We're spread so [3:11] thin. How are we going to do all the small things  answering all these small questions doing all this [3:20] kind of navigation and being able to to meet  that two also connected to staff time but also [3:26] for sustainability the organization we can't be up  24/7 365 so is there something that can do that uh [3:36] third um was that a lot of the guys especially the  new guys remember they uh are trying to turn uh a [3:44] new page, a new chapter in their life. Almost all  of them are just uh out of incarceration, working [3:51] and living in sober living. They have a lot of  questions. Some of the big questions, 100% needs [3:57] to go to staff, but most of the questions are the  same questions over and over again. And finally, [4:03] they want to make sure that whatever they're doing  doesn't replace the human touch that makes their [4:09] organization so great and so successful. So, at  first I said, "Oh my gosh, I know this solution [4:17] chatbot. We can build this amazing infrastructure,  everything." And I went back and pitched that idea [4:23] 13 months ago, and this is what they heard, right?  They said, "Oh my gosh, too much, too fast." [4:31] uh that is an AI takeover. We don't trust tech. We  don't trust government. We don't trust in general. [4:38] And I said, wait, wait, wait, wait, hold on. I'm  not going to try and academic too much of this [4:43] on the background, but let's say in general, I'm  going to this is a major oversimplification. Uh [4:50] if you're a data science person, uh you can 100%  correct me after. I like to think of chat bots in [4:57] three separate big architecture sides. On the far  end, you have heavily personalized. You're going [5:04] to see that a lot. If you're in like the medical  side and you're working through your uh electronic [5:09] medical health record, you can have extremely  personalized warm handoffs. That's exactly [5:14] the type of heavy integration that they were  initially thinking about, but not the one that [5:19] would be successful. That was to be honest kind  of the integration that I had originally thought [5:25] would be something that's heavily personalized  login uh helping somebody customize in that area. [5:32] But what I found out at the beginning was that  would have been incorrect. Um so then you also [5:37] have semi-personalized where you can have it where  somebody can choose to log in and maybe some warm [5:42] handoffs a little bit. But what we landed on was  the successful framework that is currently working [5:49] today. Um so it's completely anonymous right from  that that last slide of being too much too fast. [5:56] Um and it's purely based on conversational and  because we took a step back and and looked at [6:02] that on the implementation science side on how to  actually make it work customize that organization [6:09] um it's something that is still working today.  So if there are kind of four major elements that [6:15] I'll touch on really quickly and then I'll wrap  it up here at the end. So the first one's kind [6:20] of obvious. It's that standard knowledge base.  What what is everything in a knowledge base if [6:28] you were building your chatbot that is specific so  that it it knows exactly what it needs to know? Um [6:36] and I'll go over kind of the the main areas. The  second one um was uh the second one with that kind [6:43] of question mark and then the voice down there at  the end. Those work hand in hand. So if the voice [6:48] down there at the end, that's its persona, but  also its ability to identify safety off-ramping [6:57] when does Jesse need to know when somebody  is talking about cravings or they're talking [7:02] about hurting themselves or others. And then also  making sure in that kind of question mark box if [7:10] you're looking at that what are what's the actual  language that that specific group in northeastern [7:17] Tennessee uses. Um and after three rounds  of implementing with staff and pilot testing [7:25] um we now have uh working Jesse AI that's uh  working throughout the organization. So feel free [7:31] log on use Jesse if you want. Uh, it's completely  anonymous, as I said. It doesn't track your name. [7:37] Uh, uh, you can use Jesse as a chatbot. Uh, it  has 247 access, right? One of the things that they [7:43] needed. Um, they're a 12stepbased program. Um,  that's the type of recovery that this nonprofit [7:49] does. Um, a key area is that it also has 398  verified resources for that area. So, if you say, [7:58] "What are the Alcoholic Anonymous meetings in  Vegas?" It's probably not going to know. But if [8:03] you typed in Johnson City for example, it will be  able to pull those up. It also has things that the [8:10] guys always had for little questions which are  house rules, policies and most important as I [8:16] had said before it has different safety offramps.  Um and the background infrastructure um is built [8:25] on a ra library. Um, and it uses a platform  instead of myself doing that that development, [8:34] which was another key area because I want this  to not be, hey, a tech guy came to help out a [8:41] nonprofit and now it doesn't go anywhere. It needs  to be something that as I transition my focus into [8:47] other projects, they can implement and update  themselves. All right. So, what does the data [8:53] say? Um so this project is connected to my final  dissertation connected to my doctorate. Um and we [9:00] have three different main data points. Again this  is very still preliminary as I'm still working on [9:06] it. Uh but let's first look at some usage data. So  in this study we have 169 total residents. of that [9:14] um 125 chose to participate in in the study. Over  the 7-day kind of look at their implementation, [9:23] we had about 1,500 total questions or queries.  Um and of those about 700 were conversations. So [9:32] what that tells me is uh of those conversations it  took they were kind of chatting in one stream of [9:39] thought about two questions for each stream before  they found the information that they wanted or [9:47] didn't. But really instead of talking about data  I wanted to talk about two outstanding gentlemen. [9:54] This is Justin and Danny, two guys in the program,  and they were kind enough to share how they use [10:00] Jesse Aai. Um, so, uh, first, uh, Justin, he  used Jesse Aai to find 12step recovery meetings, [10:09] resource navigation with, uh, hyperlin, uh, GPS.  Um, so he could just click on it and know exactly [10:17] where he needed to go. It had house rules and he  could look up things for policies. Now, Danny, [10:24] Danny's been in the program and successfully in  recovery for about 2 years. Um, and so he was not [10:31] only looking at it to help with his own cravings,  but also cravings of people within his house [10:37] as he's helping them with mentorship, finding  resources, seeking help, and sponsorship planning, [10:43] which in uh the 12step world is essentially  helping others um after you've helped yourself. [10:50] Next in our kind of data for triangulation u  was an individualbased survey. So after the [10:57] guys had been using the app for an entire week  um I used the mobile application rating scale [11:05] uh which is a really simple survey um that looks  at four areas engagement functionality aesthetics [11:12] and information. Essentially that's just a bunch  of numbers. If it makes your eyes go cross, uh, [11:17] feel free to keep using Jesse AI to to chat about  something. Um, but really what that looked at, [11:23] if one is completely inadequate in that area,  five is excellent in general, um, for the overall [11:32] quality of the way that the guys independently  and anonymously rated the app. They rated it as [11:37] good. So, still not where I want it, um, but it's  at least working. They're using it. um and they [11:44] rate that it's good. Um myself as a qualitative  researcher though, this is what gets me excited. [11:50] I love focus groups. You want to run focus  groups? Let's go. Uh so we we looked at um three [11:57] main areas in buckets. First, uh what worked,  what didn't work, and what can be improved. [12:05] So the guys really connected with how fast it  was, the meeting locations, its response time, [12:13] um basic needs, those types of things. Areas that  didn't work, right? And and part of my the end of [12:19] my preliminary analysis that I'm still working  on is to categorize those into different areas. [12:25] But you had things like coverage gaps, needing to  update different resources. Um [snorts] there was [12:32] some technical instabilities when I first launched  it. There was so much usage that it actually [12:37] crashed. Uh and we had to like I had to like  scramble mid mid launch to get it up again. Um and [12:44] then working more on that voice um when there was  some feedback on how homogeneous each of this each [12:51] of the responses were. [snorts] Um and then also  things that can be improved. So increasing the [12:57] knowledge base that it pulls from um making sure  that there's content and a lot on that voice and [13:03] user experience. So when I start to really look at  triangulating a lot of this data where I'm looking [13:10] at pulling this together. Um again this is still  a very preliminary kind of those those three main [13:16] insights is what I had enough uh capacity to to  effectively triangulate now which is look they're [13:24] actually using the app. Um it seems to be really  good for the new guy coming into the program. That [13:32] was that was kind of an an undercurrent that was  pulled out of a lot of those focus groups was some [13:37] of the the older guys who had been in the program  for a year or two said, "Man, this is cool, but I [13:43] would have loved this being new. So, I'm excited  to see where this goes in over the next couple [13:48] months." And then finally on those individual  surveys that they found that it was actually [13:52] working pretty well. So again back to this kind  of base structure and this is the implementation [14:02] science part I guess you could say. So again  there's that that knowledge base and when I was [14:09] at and looking at particularly those focus groups  um the the first kind of bits of feedback really [14:17] fall into this knowledge base. So ensuring that  the resources that are in there, ensuring that the [14:25] policy book and the policy area can be updated  and then also connecting to that bottom right [14:32] uh graphic of the platform. So, ensuring that when  there's a new policy update within the nonprofit, [14:41] because I care most about implementation and  getting this in there, continuing to train the [14:47] nonprofit staff who are not tech people, right,  to be able to really easily update um essentially [14:54] this knowledge base, this RA library, um so that  they're able to have instant updates for Jesse. [15:03] But really um I think that the success or  preliminary success really happened with the [15:10] interplay between uh both the question mark and  the speaker bubble which is again that kind of [15:19] implementation science part maybe the soft skills  part the the qualitative part to make that uh that [15:28] kind of question mark area um after essentially  launching Jesse me about uh maybe 9 months ago [15:37] and launching it with staff and it kind of fell  on its face. The reason it did that um from all [15:42] my user feedback was really that it didn't have  like the voice and persona because it didn't have [15:49] the voice persona and characteristics and kind of  sayings of the internal organization. um it didn't [15:56] work. And so that led to a comprehensive set of  interviews that I did with both of the founders, [16:04] all eight of their staff and ensuring that uh then  all of that was kind of fed into being another set [16:13] of like knowledge base that then when the like  the voice of Jesse or the text voice of Jesse [16:19] grabs information, it's kind of grabbing that  voice filter first and then it's grabbing the [16:26] knowledge base before it gives out the answer.  So, I want to give a quick thank you, of course, [16:34] to the recovery resources staff um because  without them, none of this could happen. Uh they [16:40] do absolutely amazing work um within northeastern  Tennessee. Um and I'm excited uh to see the rest [16:48] of our our panelists hear everything that they  have to do and share. Um, and if you have any [16:54] ideas, you want to connect, um, feel free, please  connect. Uh, if you want to do any research stuff, [17:00] uh, and academic stuff, I would love to, uh, find  a way to do that. And if you're a missiondriven [17:05] organization, um, and you want to find a way to  partner, you can connect through there, too. So, [17:10] with that, thank you and good morning,  everyone. It's great to see everyone.