DrunR Restricts LLMs to Messaging to Curb Medical Decision Errors

Beyond the LLM: AI That Understands Before It Speaks With DrunR — Ai4

Ai4

Healthcare AI systems risk dangerous hallucinations by asking models what patients should do rather than evaluating their physiological state first, requiring deterministic rules to drive recommendations while language models serve solely as translators.

AI-assisted summary of Beyond the LLM: AI That Understands Before It Speaks With DrunR — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background Published by Ai4 on YouTube on August 19, 2026, this talk features Yaya Ali, Founder and CEO of DrunR. The presentation outlines an architectural shift in healthcare AI away from generic prompt-based generation toward an explainable system that interprets a user's physiological state before generating recommendations.

## TL;DR Yaya Ali addresses the 'invisible middle' of healthcare—the unmonitored 90-day intervals between doctor visits where patients make critical daily health decisions without clinical guidance. He explains why standard prompt-driven LLM architectures fail in high-stakes environments due to hallucination and lack of auditability. To solve this, DrunR introduced a five-step state-based framework that applies deterministic clinical rules and multimodal signals to define a patient's state before using an LLM solely as a multilingual communication agent. This approach ensures traceable 'verification receipts' and context-aware behavioral support for patients and providers alike.

## Atmosphere & style Ali opens with lighthearted, self-deprecating humor about AI hallucinations before settling into a pragmatic and methodically structured keynote. The presentation feels deeply grounded in practical engineering realities and patient safety, using relatable personal anecdotes—such as heat exhaustion on the Las Vegas Strip—to make abstract architectural trade-offs immediately clear. The overall tone is earnest, precise, and constructive, emphasizing responsible AI design over uncritical industry hype.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) The DrunR team centers their morning huddle around two primary questions: how to deliver an effective digital health product and whether AI recommendations can be trusted when patient health and lives are on the line. - [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) Healthcare suffers from an 'invisible middle'—up to 90 days between 30-minute doctor visits where patients manage complex medication, nutrition, and symptom decisions entirely on their own. - [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) DrunR serves as a behavioral intelligence layer in the invisible middle, offering patient guidance via wearables and food scanning while flagging aggregated risk signals to providers on a dashboard. - [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) DrunR's initial architecture used a 4,000-token prompt passed to a frontier LLM, which failed because black-box reasoning could not generate auditable verification receipts for physiological recommendations. - [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) Asking an LLM 'What should this person do?' leads to misinterpreted context, whereas asking 'What state is this person in right now?' accurately determines underlying physiological needs like hydration and joint recovery. - [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) DrunR's state-based architecture operates in five steps: signal collection/normalization, clinical boundary application, user state derivation, matching to qualified recommendation tables, and LLM communication. - [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) In a state-based architecture, the LLM is restricted to being a communication agent and translator rather than a core reasoning and decision-making engine. - [09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) Separating state reasoning from language generation ensures high auditability with traceable breadcrumbs and verification receipts for every generated recommendation. - [09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) Deterministic clinical rules and guardrails prevent dangerous or generic advice, such as misaligned macronutrient targets for patients on GLP-1 medications. - [09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) The core design philosophy for high-stakes AI applications is to build systems that 'understand first, then speak second' by evaluating state before generating language.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) Morning Huddle Questions and the "Invisible Middle" of Healthcare

Yaya Ali, Founder and CEO of DrunR, opens his presentation with a lighthearted story about preparing his talk the night before. He asked an LLM to generate an opening joke for the audience. The model suggested walking onto the stage, coughing, and saying, "Sorry, I just got up from a bit of a hallucination." This humorously points directly to the central challenge of his presentation: trust in AI recommendations. Every morning during the DrunR team huddle, they address two primary questions: how to deliver the most effective version of their digital health product to users and partners, and whether they can truly trust the AI output and recommendations when patients' health and lives are on the line.

```mermaid flowchart TD accTitle: Patient Healthcare Journey and the Invisible Middle

subgraph S1["Clinical Touchpoints"] A["Initial Doctor Visit (30 mins)"] B["Follow-up Visit (Up to 90 Days Later)"] end

subgraph S2["The Invisible Middle"] C["Day-to-Day Self-Care"] D["Daily Questions: Medications, Diet, Symptoms"] E["Unmonitored Risk Accumulation"] end

A --> C C --> D D --> E E --> B ```

The core problem DrunR addresses lies in the current healthcare workflow. A patient sees a doctor for 30 minutes to establish a treatment plan, but their next follow-up appointment might not occur for up to 90 days. Ali terms this intervening period the "invisible middle"—hundreds of days and thousands of hours where patients are left entirely on their own to navigate daily decisions regarding medications, nutrition, and unexpected symptoms like nausea or appetite loss. Doctors cannot monitor patients in real time during this period, so when a follow-up visit finally occurs, time is spent investigating setbacks (e.g., unexplained weight gain, high blood pressure, fluctuating blood sugar) rather than focusing on progress.

- DrunR is designed to operate directly within the "invisible middle" as a behavioral intelligence layer for both patients and healthcare providers. - For patients, DrunR offers nutrition and activity recommendations by connecting to wearable hardware (smartwatches, Oura rings, continuous glucose monitors) and scanning surrounding food options (grocery stores, restaurants, home cooking). - For providers, DrunR aggregates aggregated risk signals—such as poor sleep patterns or irregular heart rates—and highlights them on a dashboard so healthcare teams can intervene proactively rather than waiting 90 days.

[00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) Shifting from Prompt-Based Models to State-Based AI

When DrunR began its development journey a year prior, the initial architecture relied on a standard prompt-based approach. The system gathered multimodal input signals (hardware stats, location data, current medications) and wrapped them into a single long prompt of roughly 4,000 tokens before passing it to a frontier LLM to produce a recommendation. While this yielded confident-sounding answers, it ultimately failed when the team attempted to audit outputs and generate "verification receipts" explaining why a specific recommendation was given at a particular moment. Prompt-based LLMs proved incapable of accurately understanding the patient's underlying physiological state.

```mermaid flowchart LR accTitle: Prompt-Based vs State-Based Architecture Comparison

subgraph Standard["Prompt-Based Approach (Legacy)"] P1["Data Inputs"] --> P2["4k Token Prompt"] P2 --> P3["LLM Reasoning"] P3 --> P4["Generic Output"] end

subgraph StateBased["State-Based Approach (DrunR)"] S1["Data Inputs"] --> S2["State Engine"] S2 --> S3["Rules & Boundaries"] S3 --> S4["Qualified Recs"] S4 --> S5["LLM as Communicator"] end ```

To illustrate the flaw of prompt-driven recommendations, Ali shares a personal anecdote. After walking 20,000 steps in a day and experiencing severely interrupted sleep, a standard LLM might congratulate him for being a "champion" and suggest rewarding himself with a big slice of cake. However, half of those steps were walked outside on the Las Vegas Strip in 113-degree heat while the other half were indoors, leading to extreme fatigue and joint stress rather than healthy exercise. Asking the wrong question—"What should this person do right now?"—causes LLMs to misinterpret context, whereas asking "What state is this person in right now?" allows the system to determine that the immediate need is hydration, joint recovery, and reduced heat exposure.

- The shift from prompt-based to state-based AI replaces the question "What should the person do?" with "What state is this person in right now?" - DrunR's architecture relies on five explicit processing steps: 1. Signal collection and data normalization into structured formats. 2. Application of clinical rules, guidelines, and safety boundaries. 3. Derivation of an explicit user state definition. 4. Matching the defined state to qualified options from pre-built recommendation tables. 5. Utilizing the LLM strictly as a communication agent to translate medical concepts into plain, accessible language across multiple languages.

[09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) Benefits of Auditable State-Driven AI in Healthcare

By restricting the LLM's role to communication rather than reasoning, DrunR eliminates the "black box" risks inherent in pure LLM architectures. The system maintains detailed trail "breadcrumbs" from raw inputs to final outputs, allowing engineers, clinicians, and patients to trace any recommendation back to its underlying data and clinical rule logic. If a patient asks why a specific meal or restaurant item was suggested, DrunR can instantly generate a precise verification receipt.

```mermaid flowchart TD accTitle: DrunR Five-Step State Architecture Flow

subgraph DataPrep["1. Signal Collection"] A["Wearable Hardware"] B["Location & Context"] C["Medication Data"] A & B & C --> D["Data Normalization"] end

subgraph StateEngine["2. State Analysis & Logic"] D --> E["Apply Rules & Guidelines"] E --> F["Define Patient State"] F --> G["Filter Qualified Recs Table"] end

subgraph Communication["3. Output Generation"] G --> H["LLM Translation Agent"] H --> I["Patient Message & Audit Receipt"] end ```

This structural separation ensures that clinical guardrails and boundaries remain deterministic, preventing generic or unsafe responses—such as standard macro advice for patients on GLP-1 medications—and replacing them with personalized, state-aware guidance. Provider teams gain full visibility into patient health trends across the 90-day interval between clinical visits, eliminating dangerous care blind spots. Ali concludes with a key takeaway for developers across all industries: build systems that "understand first, then speak second" by defining user states and boundaries before invoking LLMs for communication.

- Auditable "breadcrumbs" allow DrunR to generate verification receipts for every output shown to patients and physicians. - The LLM acts solely as a communicator capable of translating complex medical terms into clear, multi-lingual daily messages. - For business inquiries and developer outreach, Yaya Ali provided the contact email `support@d-runner.com`.

## Notable quotes > Most AI applications follow the same pattern: collect data, send it to an LLM, and generate an answer. I believe that approach breaks down when decisions have real-world consequences, especially in healthcare. — Yaya Ali

> [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) Sorry, I just got up from a bit of a hallucination. — Yaya Ali

> [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) invisible middle — Yaya Ali

> [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) verification receipts — Yaya Ali

> [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) What should this person do right now? — Yaya Ali

> [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) What state is this person in right now? — Yaya Ali

> [00:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=45s) What should the person do? — Yaya Ali

> [09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) black box — Yaya Ali

> [09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) breadcrumbs — Yaya Ali

> [09:45](https://www.youtube.com/watch?v=C4WMCV5e94k&t=585s) understand first, then speak second — Yaya Ali

> designed to support better decisions, not just better conversations. — Yaya Ali

## Mentioned Yaya Ali, DrunR, Ai4, Oura, Las Vegas Strip, GLP-1

## Video description Yaya Ali - Founder & CEO, DrunR

Most AI applications follow the same pattern: collect data, send it to an LLM, and generate an answer. I believe that approach breaks down when decisions have real-world consequences, especially in healthcare. In this talk, I’ll share how building DrunR led us to a different architecture, one that transforms multimodal data into an explainable understanding of a person’s state before an LLM ever generates a recommendation. You’ll leave with a practical framework for building AI systems that are more trustworthy, more personalized, and designed to support better decisions, not just better conversations.

[Source](https://www.youtube.com/watch?v=C4WMCV5e94k)

中文简报

DrunR限制大语言模型仅用于信息传递以遏制医疗决策错误医疗AI系统若直接询问模型患者该怎么做、而非先评估其生理状态，就会面临危险的幻觉风险；因此需要由确定性规则来驱动建议，而语言模型仅用作翻译。 Yaya Ali探讨了医疗领域的“隐形中间地带”——即两次就诊之间长达90天的无监测间隔期，在此期间患者在缺乏临床指导的情况下做出一系列关键的日常健康决策。他解释了标准提示词驱动的LLM架构为何因幻觉和缺乏可审计性而在高风险环境中失效。为解决这一问题，DrunR引入了一套基于状态的五步框架，先运用确定性临床规则和多模态信号界定患者状态，随后仅将LLM用作多语言沟通智能体。该方法确保了可追溯的“核验凭证”，并为患者和医疗提供者提供具备情境感知的行为支持。 DrunR 聚焦于两次就诊间长达 90 天的“隐形中间期”（invisible middle），通过整合可穿戴设备与食物扫描数据为患者提供日常行为指导，并在仪表板上向医生汇总风险信号。 (0:04) 团队早期采用将 4,000 token 提示词直接交由前沿 LLM 生成建议的方案宣告失败，因其黑盒推理无法为关乎生命健康的生理建议提供可审计的验证依据。 (0:45) 架构核心转变在于将提问从“这个人应该做什么”改为“这个人现在处于什么状态”，从而准确识别补水、关节恢复等潜在生理需求，避免误判情境。 (0:45) DrunR 的五步状态架构将 LLM 严格限制为沟通与翻译代理而非决策引擎，确保所有输出均符合确定性的临床边界，并具备可追溯的验证凭证。 (9:45)

Reading summary

DrunR Founder and CEO Yaya Ali argues that digital healthcare tools must abandon pure prompt-based LLM architectures in favor of state-based systems that resolve patient physiology before generating recommendations. Addressing the 'Invisible Middle' in Healthcare Workflows (0:04) Ali describes the 'invisible middle' as the extended gap—often lasting up to 90 days—between a patient's standard 30-minute doctor visits. During these hundreds of days, patients manage diet, medication, and sudden side effects such as nausea or appetite changes without clinical oversight. When follow-up appointments finally happen, clinicians must spend consultation time investigating setbacks like irregular blood pressure or fluctuating blood sugar rather than tracking health progress. DrunR is built to function within this monitoring void as a behavioral intelligence layer for patients and medical providers. The system ingests inputs from wearables such as smartwatches, Oura rings, and continuous glucose monitors, alongside local food data from grocery stores and restaurants. For medical teams, aggregated risk metrics like irregular heart rates or poor sleep patterns are compiled onto a dashboard to enable proactive clinical intervention before the next routine visit. Limitations of Prompt-Driven LLM Architectures (0:45) During its first year of development, DrunR used a conventional prompt-based design that compressed hardware metrics, medications, and contextual data into roughly 4,000-token prompts fed directly to frontier LLMs. While this produced articulate outputs, it failed whenever the team tried to audit decisions or explain why a particular suggestion was made. Ali asserts that asking an LLM 'What should this person do right now?' causes models to misinterpret physiological context because they cannot accurately deduce the user's underlying physical state. To illustrate this, Ali cited an instance where he accumulated 20,000 steps while enduring interrupted sleep, with half those steps taken in 113-degree heat on the Las Vegas Strip. A standard LLM evaluated the metric naively, praising him as a 'champion' and suggesting a slice of cake as a reward. In reality, the physical strain called for joint recovery, hydration, and heat mitigation rather than celebratory calories. State-Based Architecture and Deterministic Clinical Logic (9:45) To prevent black-box reasoning, DrunR shifted its core question to 'What state is this person in right now?' The resulting architecture uses a five-step pipeline: collecting and normalizing multimodal signals, applying deterministic clinical rules and boundaries, explicitly defining the patient's physiological state, filtering appropriate interventions from pre-built recommendation tables, and deploying the LLM strictly as a natural-language communicator and multi-lingual translator. Restricting the LLM to communication rather than clinical reasoning creates an auditable trail of 'breadcrumbs.' Patients and clinicians can trace any recommendation directly back to underlying sensor data and medical guidelines through an instant verification receipt. According to Ali, this separation protects patients from generic advice—such as standard macronutrient plans inappropriate for individuals on GLP-1 medications—while maintaining safe clinical boundaries. Ali advises AI developers to design architectures that understand context and establish boundaries before invoking language models. While DrunR demonstrates this framework for behavioral health monitoring, the presentation focuses on architectural structure and does not provide clinical trial data or performance benchmarks across broader patient populations.

精炼阅读摘要

DrunR 创始人兼首席执行官 Yaya Ali 指出，传统医疗在长达 90 天的就诊间隔期内存在监测盲区，而依赖大语言模型直接决策存在严重的“黑盒”与幻觉风险；因此，医疗 AI 必须转向状态驱动架构，在调用模型发声前先确定患者的生理状态与临床安全边界。就诊间隔期与日常自我护理盲区 (0:04) 在现行医疗服务流程中，患者通常仅能与医生进行约 30 分钟的面诊以制定治疗方案，而下一次复诊往往要等待长达 90 天。Ali 将这段涵盖数百天与数千小时的间隙称为“无形中间期”（invisible middle）。在此期间，患者必须独自面对复杂的日常健康决策，包括用药管理、营养摄入以及恶心或食欲不振等突发症状。由于缺乏实时监测机制，当复诊最终到来时，医护团队往往被迫耗费大量时间排查体重反弹、血压偏高或血糖波动等病情倒退的原因，而无法专注于长期治疗进展。针对该痛点，DrunR 被设计为一个介于患者与医疗服务提供方之间的行为智能层。在患者端，该系统接入智能手表、Oura 智能戒指以及连续血糖监测仪等可穿戴设备，并结合杂货店、餐馆和家庭烹饪等周边食物环境，提供个性化的营养与活动建议；在医生端，系统持续汇集睡眠紊乱或心率异常等风险信号并在仪表板呈现，使医疗团队能够主动发起干预，消除长达 90 天的照护真空。提示词驱动决策的黑盒缺陷 (0:45) DrunR 在研发初期曾采用基于提示词的标准架构：系统将可穿戴设备数据、位置信息和用药记录整合为一个约 4000 token 的长提示词，直接交由前沿大语言模型生成建议。然而，这一模式虽然能给出看似确定自信的回答，但在系统审计与临床归因时却彻底失效。团队发现，大语言模型无法真正理解患者底层的生理状态，其决策过程缺乏透明度，无法生成能够解释具体建议生成依据的“验证凭证”。 Ali 以自身经历说明了该模式的系统性误判：他在华氏 113 度（约 45 摄氏度）的拉斯维加斯大道户外与室内合计行走了 20000 步且夜间睡眠严重受阻，传统大语言模型却将其识别为运动达人并建议吃大块蛋糕作为奖励。Ali 指出，向模型询问“用户现在应该做什么”会导致上下文被曲解，掩盖了关节劳损、极度疲劳和中暑脱水的真实危险；只有将核心问题转变为“患者当前处于何种生理状态”，系统才能准确识别出补水、关节恢复与避免高温暴晒的紧迫需求。五步状态引擎构建确定性边界 (0:45) 为从根本上解决逻辑黑盒问题，DrunR 确立了先定义状态、再筛选方案、最后由大模型转译的五步处理流程。第一步是多模态信号采集与数据规范化，将各类硬件与环境指标转化为结构化数据；第二步是引入临床指南、专业医学规则与安全边界条件；第三步是运算推导并明确定义出患者当前的具体生理状态，替代传统直接向大模型发问的模式。在完成状态定义后，第四步是将该状态与预先构建的合格建议库进行规则匹配，确保所有输出均处于严格的临床许可范围内；第五步则将大语言模型严格限定为“沟通转译智能体”，仅负责将复杂的医学概念与方案转化为多语言、通俗易懂的日常表达。这种架构剥离了大模型的推理决策权，确保临床防线始终保持确定性，有效防止了在应对 GLP-1 类药物患者等特殊群体时生成泛化或危险的膳食建议。全流程验证凭证与临床可追溯性 (9:45) 将大语言模型降级为纯沟通媒介后，DrunR 消除了端到端生成式模型固有的不可预测性。系统在原始输入数据与最终建议之间建立了完整的审计追踪链（breadcrumbs），使工程师、临床医生及患者本人能够清晰复盘每一条健康建议背后的底层数据指标与临床规则逻辑。例如，当患者对特定餐食或推荐菜品提出疑问时，系统能够即时出具详尽的验证凭证。这种全流程可审计性不仅帮助医疗团队在 90 天的随访周期内完全掌握患者的健康变化轨迹，也使数字疗法在面临医疗责任审查时具备坚实的确定性支撑。Ali 据此向行业开发者提出建议：在设计涉及生命健康的智能化系统时，必须建立“先理解状态、后开口表达”的架构规范，将规则审计前置，而非让生成式大模型承担超出其能力的临床逻辑推理。 Ali 强调健康监测必须以确定性状态定义取代黑盒推理，但该架构在不同慢性病场景下的规则拓展成本，以及在更广泛医疗生态中对接异构数据的通用性，仍有待长期的临床实证检验。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2026 年 8 月 19 日发布在 YouTube 平台。内容由 DrunR 创始人兼 CEO Yaya Ali 主讲，探讨了在关乎患者健康的医疗决策场景中传统“收集数据直接输入 LLM”模式的局限性，并分享了 DrunR 如何通过多模态数据在模型输出前精准理解用户状态的系统架构。 ## TL;DR 演讲指出了患者在两次就诊间隔长达 90 天的“隐形中间期”脱节痛点，以及直接用大 Prompt 驱动 LLM 决策所面临的黑盒不可解释与幻觉风险。为此，DrunR 提出了从“Prompt 范式”向“State 状态范式”的核心转向，建立起涵盖数据标准化、医学规则约束与状态推荐库匹配的五步闭环架构。在该系统中，LLM 不再承担医学逻辑决策，而是退回到其最擅长的通信与语言转译层，从而为患者和医生提供具备完整可追溯审计凭据的高可信个性化干预方案。 ## Atmosphere & style 演讲整体务实、严谨且极具启发性。主讲人以幽默的“AI 幻觉笑话”开篇破题，随后迅速切入关乎患者生命的严肃医疗信任议题，并在分享中穿插自己在拉斯维加斯高温下暴走的亲身案例，生动揭示了纯大模型盲目推理与精准状态感知之间的强烈反差。整场演讲逻辑层层递进，既有对行业黑盒痛点的深刻反思，又有清晰的技术解法，展现出强烈的工程落地感与专业信服力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) 医疗健康场景下 AI 的首要挑战是建立信任：当决策关乎患者生命健康时，必须杜绝模型幻觉，确保输出结果高度安全可靠。 - [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) 传统诊疗流程存在长达 90 天的“隐形中间期”，期间患者脱离专业照护，极易导致复诊时医生的大部分时间退化为倒查病情倒退的追责与调查。 - [03:30](https://www.youtube.com/watch?v=C4WMCV5e94k&t=210s) DrunR 提出构建“行为智能层”，通过整合穿戴设备生理信号与周边环境营养数据，无缝填补两次门诊之间的日常健康管理空白。 - [03:30](https://www.youtube.com/watch?v=C4WMCV5e94k&t=210s) 将系统定位为医疗团队的“风险聚合器”，将睡眠或心率等异常指标实时呈现给主治医生，变被动等待复诊为主动即时干预。 - [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) 单纯依赖 Prompt 的端到端黑盒架构存在致命缺陷：把海量原始数据打包丢给 LLM 会失去可解释性，无法提供严谨的因果审计凭据。 - [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) AI 范式应从“直接询问该做什么”转向以“状态”（State）为核心：系统必须后退一步，优先明确患者当下的生理与环境真实状态。 - [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) 极端情境印证了状态建模的重要性：在酷暑下的 20,000 步属于严重疲劳与脱水状态，不能简单依步数数据给出鼓励性甚至有害的盲目建议。 - [08:44](https://www.youtube.com/watch?v=C4WMCV5e94k&t=524s) 落地五步闭环架构（收集、清洗标准化、规则判定、推荐匹配、LLM 通信），为医疗决策建立如“面包屑轨迹”般可全程追溯的验证凭据。 - [08:44](https://www.youtube.com/watch?v=C4WMCV5e94k&t=524s) 对大模型进行职能降级：将 LLM 从核心医学逻辑推理者转变为专职的“沟通转译代理”，负责将严谨推导出的专业方案转化为患者易懂的白话。 - [10:55](https://www.youtube.com/watch?v=C4WMCV5e94k&t=655s) 确立“先理解、后表达”（Understand first, speak second）的系统构建准则，先进行状态感知与业务边界约束，最后才由语言模型负责呈现输出。 - [10:55](https://www.youtube.com/watch?v=C4WMCV5e94k&t=655s) 消除长周期空档期的就医盲区，持续追踪患者行为与生理状态，使医疗机构在 90 天脱节期内依然具备对潜在健康隐患的敏锐洞察力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) 医疗健康场景下的 AI 信任危机与“隐形中间期” DrunR 创始人兼 CEO Yaya Ali 在开场时分享了一个幽默插曲：前一天晚上他让大语言模型（LLM）帮他构思一个开场笑话，LLM 建议他直接走上台、一边咳嗽一边说：“抱歉，我刚从幻觉（Hallucination）里醒过来。” 这个笑话引出了他在医疗 AI 领域最核心的思考——每天早会团队都要面对两个问题：第一，如何为用户和合作伙伴交付更高效的产品；第二，也是最关键的，**当决策关乎患者的健康乃至生命时，我们真的能信任 AI 模型的输出和推荐吗？** 在传统的医疗就诊流程中，患者通常在身体不适时去医院见医生约 30 分钟，确定治疗目标和方案，而下一次复诊可能要等长达 90 天。在这长达数十天甚至数千小时的时间跨度中，存在一个巨大的空白地带，即**“隐形中间期”（Invisible Middle）**。患者在此期间只能独自面对日常疑问：“我今天吃药了吗？”、“我该吃什么？”、“为什么我感到恶心或者食欲不振？”而医生无法做到实时跟进。当 90 天后再次见面，如果患者体重增加、血压飙升、心率或血糖异常，医生宝贵的 20 分钟就诊时间就会从“讨论治疗进展与下一步方案”彻底沦为一场“排查病情为何倒退的追责与调查”。 - 患者就诊周期：初次门诊约 30 分钟 $\rightarrow$ 长达 90 天的脱节期（隐形中间期） $\rightarrow$ 仅约 20 分钟的复诊调查。 - 隐形中间期的痛点：缺乏实时干预，导致健康隐患累积，破坏原本的治疗依从性与效果。 --- ### [03:30](https://www.youtube.com/watch?v=C4WMCV5e94k&t=210s) DrunR 的定位：构建行为智能层以填补空白 DrunR 的诞生正是为了切入这一“隐形中间期”，打造一个介于患者与医疗团队之间的**行为智能层（Behavior Intelligence Layer）**。系统通过连接患者的智能穿戴设备（如智能手表、Oura Ring 智能指环、连续血糖监测仪 CGM 等）获取生理信号，并扫描周边环境中的营养选项（包括杂货店、餐馆外卖及家庭烹饪），为患者提供个性化的饮食与身体活动建议。对于医疗机构和医生端，DrunR 扮演着风险聚合器的角色。如果系统在两次就诊之间监测到患者睡眠极差或心率剧烈波动，它会将这些风险特征提炼出来并呈现在医生的实时监控仪表盘上。这样医生无需被动等待数月后的复诊，而是能在异常发生的当下主动介入，进行实时干预。 ```mermaid flowchart LR accTitle: DrunR 行为智能层工作流架构 subgraph S1["数据感知"] A["智能穿戴设备 (手表 / Oura / CGM)"] B["环境与饮食信号 (超市 / 餐厅 / 家常菜)"] end subgraph S2["DrunR 智能处理"] C["行为智能层 (聚合与风险计算)"] end subgraph S3["双向协同反馈"] D["患者端 (实时日常行动建议)"] E["医生仪表盘 (实时风险预警干预)"] end A --> C B --> C C --> D C --> E ``` --- ### [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) 从 Prompt 范式到 State 范式：AI 范式的核心转向团队最初探索 LLM 应用时，遵循了行业常见的 Prompt 范式：将可穿戴设备数据、地理位置、当前用药及环境参数等各种单点数据全部打包成一个大约 4,000 Token 的 Prompt 喂给大模型，让 LLM 直接给出推荐。虽然大模型的生成语气非常自信，但团队很快遇到了不可逾越的瓶颈——**无法审计（Auditing）与归因**。医生和患者会质疑：“你为什么在这一刻给我这个具体的推荐？”以 Prompt 为核心的模型根本无法提供清晰的验证凭据（Verification Receipt）。 ```mermaid flowchart TD accTitle: 从基于 Prompt 到基于 State 的范式转变 subgraph OLD["过去：Prompt 驱动（黑盒推理）"] O1["收集多模态原始信号"] --> O2["拼装 4000 Token Prompt"] O2 --> O3["LLM 直接推理决策 (提问: 该做什么?)"] O3 --> O4["黑盒输出: 缺乏可解释性与审计凭据"] end subgraph NEW["现在：State 驱动（规则 + 状态机）"] N1["数据归一化与特征处理"] --> N2["规则引擎判定患者真实状态 (提问: 处于什么状态?)"] N2 --> N3["匹配合规推荐表库"] N3 --> N4["LLM 仅作通信与语言翻译"] N4 --> N5["全链路面包屑轨迹与可审计凭证"] end ``` Yaya Ali 意识到，直接向 LLM 询问“患者现在应该做什么？”从根本上就问错了问题。系统必须后退一步，将核心问题转变为：**“这个人现在处于什么状态（State）？”** Yaya Ali 用自身经历举例：他前一晚走了 20,000 步，但夜间睡眠破碎，每两小时醒来一次。如果将这些简单数据丢给 ChatGPT 等主流 LLM，模型大概率会欢呼：“太棒了！你完成了 20,000 步，真是个冠军，快奖励自己一块大蛋糕吧！”但事实真相是：这 20,000 步有一半是在拉斯维加斯 113°F（约 45°C）的高温户外行走的，另一半是在室内展厅穿梭。他并没有在“积极健身”，而是处于严重的**极度疲劳与脱水状态**，这也是他无法安睡的根源。基于“状态（State）”的系统生成的正确指导应为： - 晨起首要任务：立即补水（Hydration）。 - 运动限制：减少室内步行，严禁暴露在高温室外。 - 康复营养：推荐针对关节疲劳和肌体恢复的特定营养搭配。 --- ### [08:44](https://www.youtube.com/watch?v=C4WMCV5e94k&t=524s) 五步状态驱动架构与可解释性落地为了彻底摆脱 LLM“弱模型做过载决策”带来的黑盒风险，DrunR 构建了五步闭环架构： 1. **收集信号（Collect Signals）**：聚合多模态健康与环境数据。 2. **数据标准化（Normalize Data）**：将非结构化信号清洗为结构化数据。 3. **规则与指南应用（Apply Rules & Guidelines）**：通过医学准则明确界定患者当前所处的精准状态。 4. **状态匹配推荐库（State-to-Recommendation Matching）**：从预先验证的医学推荐表中筛选出符合该状态的合格方案。 5. **LLM 专职通信代理（LLM as Communication Agent）**：将 LLM 从不擅长的“医学逻辑推理者”降级为“表达与沟通工具”。 LLM 最大的优势在于支持多语言且擅长将复杂的医学术语转化为易懂的大白话（Plain English）。在这一全新架构下，系统具备了完整的**可追溯审计凭据（Verification Receipt）**，如同一条清晰的“面包屑轨迹”（Breadcrumbs），能让医生与患者随时追溯到每一个推荐动作背后的真实依据与逻辑链条。 - **消除泛化套话**：告别类似向使用 GLP-1 减重药患者简单回复“多吃蛋白、少吃碳水”的无效泛化建议。 - **强化防护栏（Guardrails）**：核心推荐来自严格的医学规则库，而非 LLM 的不可控幻觉。 - **深度个体化（Hyper-Personalization）**：每一次干预均针对患者当前的生理极限、环境温度与疲劳状态精准生成。 ### [10:55](https://www.youtube.com/watch?v=C4WMCV5e94k&t=655s) 先理解后表达：构建以状态感知为核心的 AI 系统在患者两次常规就医长达 90 天的间隔期内，传统医疗系统往往存在严重的监控盲区。Yaya Ali 强调，系统设计的关键在于确保主治医生在这漫长的 90 天空档期中依然保有清晰的洞察力与可见性（visibility），防止患者脱离照护视线，并主动挖掘和排查潜藏的不良健康隐患。面对全场观众，Yaya Ali 提炼出他在构建 DrunR 过程中最重要的核心建议：“尽量让你的模型先理解，再表达”（"try to make your model understand first then speak second"）。无论开发者身处零售、医疗还是其他任何业务场景，都不应让大语言模型从原始数据直接盲目输出建议，而是应当首先针对最终用户（无论是患者还是消费者）建立一套精确反映其实际处境的“状态”（State）建模，摸清他们的真实需求，随后施加严格的专业与业务边界控制，最后再把舞台交还给大模型——让 LLM 发挥其最擅长的本领，即充当沟通与语言转译层，将严谨推导出的状态和结论清晰易懂地传达给患者。 ```mermaid flowchart TD accTitle: 先理解后表达的系统架构 subgraph Understanding["第一阶段：状态理解"] A["用户/患者多维数据"] --> B["状态建模 (State Modeling)"] B --> C["挖掘就诊盲区中的真实需求与隐患"] end subgraph Boundary["第二阶段：边界约束"] C --> D["施加业务与医疗规则边界 (Boundaries)"] end subgraph Speaking["第三阶段：语言呈现"] D --> E["LLM 发挥沟通与转译长处"] E --> F["向患者输出高可信建议"] end ``` - **消除 90 天就医盲区**：通过在患者两次就诊之间的 90 天空隙期持续跟踪生理与行为状态，及时揭示未被察觉的不健康指标。 - **架构范式转变**：拒绝“收集数据直接丢给 LLM”的黑盒模式，倡导“先建模状态、再设定边界、最后由 LLM 负责表达”的递进式架构。 - **联系与对接渠道**：演讲最后，Yaya Ali 感谢现场观众，并表示参会者可以通过大会专属 App 查找其个人主页，或直接通过官方支持邮箱 `support@d-runner.com` 与 DrunR 团队取得联系。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) 抱歉，我刚从幻觉（Hallucination）里醒过来。 — Yaya Ali > [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) 当决策关乎患者的健康乃至生命时，我们真的能信任 AI 模型的输出和推荐吗？ — Yaya Ali > [00:04](https://www.youtube.com/watch?v=C4WMCV5e94k&t=4s) 讨论治疗进展与下一步方案彻底沦为一场排查病情为何倒退的追责与调查。 — Yaya Ali > [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) 你为什么在这一刻给我这个具体的推荐？ — Yaya Ali > [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) 这个人现在处于什么状态（State）？ — Yaya Ali > [04:53](https://www.youtube.com/watch?v=C4WMCV5e94k&t=293s) 太棒了！你完成了 20,000 步，真是个冠军，快奖励自己一块大蛋糕吧！ — Yaya Ali > [10:55](https://www.youtube.com/watch?v=C4WMCV5e94k&t=655s) try to make your model understand first then speak second — Yaya Ali > [10:55](https://www.youtube.com/watch?v=C4WMCV5e94k&t=655s) 尽量让你的模型先理解，再表达 — Yaya Ali ## Mentioned Yaya Ali, DrunR, Ai4, ChatGPT, Oura Ring, CGM (连续血糖监测仪), GLP-1 ## Video description Yaya Ali - Founder & CEO, DrunR Most AI applications follow the same pattern: collect data, send it to an LLM, and generate an answer. I believe that approach breaks down when decisions have real-world consequences, especially in healthcare. In this talk, I’ll share how building DrunR led us to a different architecture, one that transforms multimodal data into an explainable understanding of a person’s state before an LLM ever generates a recommendation. You’ll leave with a practical framework for building AI systems that are more trustworthy, more personalized, and designed to support better decisions, not just better conversations. [Source](https://www.youtube.com/watch?v=C4WMCV5e94k)

Source quotes

[0:04] Yaya Ali: 抱歉，我刚从幻觉（Hallucination）里醒过来。 [0:04] Yaya Ali: 当决策关乎患者的健康乃至生命时，我们真的能信任 AI 模型的输出和推荐吗？ [0:04] Yaya Ali: 讨论治疗进展与下一步方案彻底沦为一场排查病情为何倒退的追责与调查。 [4:53] Yaya Ali: 你为什么在这一刻给我这个具体的推荐？ [4:53] Yaya Ali: 这个人现在处于什么状态（State）？ [4:53] Yaya Ali: 太棒了！你完成了 20,000 步，真是个冠军，快奖励自己一块大蛋糕吧！ [10:55] Yaya Ali: try to make your model understand first then speak second [10:55] Yaya Ali: 尽量让你的模型先理解，再表达

Source quotes

Yaya Ali: Most AI applications follow the same pattern: collect data, send it to an LLM, and generate an answer. I believe that approach breaks down when decisions have real-world consequences, especially in healthcare. [0:04] Yaya Ali: Sorry, I just got up from a bit of a hallucination. [0:04] Yaya Ali: invisible middle [0:45] Yaya Ali: verification receipts [0:45] Yaya Ali: What should this person do right now? [0:45] Yaya Ali: What state is this person in right now? [0:45] Yaya Ali: What should the person do? [9:45] Yaya Ali: black box [9:45] Yaya Ali: breadcrumbs [9:45] Yaya Ali: understand first, then speak second Yaya Ali: designed to support better decisions, not just better conversations.

Original transcript

[0:04] Uh thank you uh for being here. Uh my name is Yaya and uh last night uh while I was getting ready for [0:11] my presentation today, uh I asked it uh like my LLM uh model uh to build me like, you know, what [0:18] is the opening joke that I should share with you. And basically, the LLM came back to me saying, [0:25] "You just walk to the stage and you start caffeine and say, 'Sorry, I just got up from a [0:30] bit hallucination.'" And uh you know, that's lead to the question today. So, we uh we are Di Runner. [0:36] I'm the founder and I help the team to build uh the uh health uh digital health uh product and uh [0:44] for our team, uh when we wake up in the morning, our morning huddle, we have two questions that you [0:49] we need to answer. The first one, how we can build and deliver like, you know, the most [0:54] uh effective and, you know, uh uh a better uh copy or you like, you know, version of our Di Runner to [1:01] our user and our partners. The second question is like, you know, would we trust uh the AI uh [1:08] output? Like, you know, it's coming from our model as a recommendation, especially when the decision [1:13] that we're sending back to the patients matter for their health or for their life. And this will [1:19] be like, you know, we would today I'll be I'll be I'll be uh focusing on the second question, [1:23] which like, you know, the trust issue that we run through. I'll be just sharing with you like our [1:27] true a true story of our internal uh solution that we built uh to mitigate that risk. And uh [1:35] before I dive in, I'm just going to give you like, you know, a hint about the environment that we we [1:40] uh we operate in. So, we operate in a health care. And basically, as a patient or an also, [1:44] so this will be the exact workflow that a patient goes through. They find out there is something [1:50] not going right with their health, there is a health problem that they want to fix. They find a [1:54] provider, they see the doctor for 30 minutes, they set up the goals, the plan for their treatment, [2:00] then uh they are there on their own. So, the next follow-up will be 90 day, up to 90 days uh [2:07] from the first initial one. And in the middle as you can see there uh is what we call the invisible [2:12] middle. So, this is when you are you are uh on your own uh dealing with the day-to-day questions [2:18] that you have to do. Uh did I take my medication today? What did I eat? What should I eat? Why I'm [2:24] feeling like, you know, different, like my nausea, is I'm losing my appetite? What's going on? How [2:28] can I help that? This invisible middle, like, you know, this is this is a count of hundred of days [2:34] over the year or thousand and thousand of hours that, you know, you are on your own trying to [2:38] make sure that the treatment plan that you set with your doctor is working right for you until [2:43] the next visit. This is for you as a patient. For your doctor, they cannot be with you in a [2:48] real time day by day trying to help you with these uh fundamental questions. The only thing they can [2:54] do is wait until the next visit. And only then, only then when they can find like the outcome, [3:00] you're gaining weight, uh your blood pressure is going high, your heart rate is not okay, [3:05] uh your blood uh sugar is is jumping around, then they start asking investigation questions. So, the [3:11] 20 minutes that you have with your doctor for the next visit, you know, flips upside down from like, [3:16] you know, what is the progress, how can we move to the next phase to become like an an investigation [3:22] to uncover the real cause why there is a setback or why our plan is not moving forward. [3:30] And for the runner, uh we are built to sit exactly within in this invisible layer. So, [3:37] for uh we're building like, you know, the the behavior intelligence layer that, you know, [3:41] for patients, we can help them with, you know, answer the day-to-day questions they have. [3:45] We do that, like, you know, by recommendation nutrition for them and the body activities. [3:50] So, basically, we connect with your guys you are if the patient is using like a smart hardware, [3:55] like smart smart watch, or like, you know, order ring, or the glucose count, [3:59] we connect with that and we read the signals. And we always scan all the nutrition uh [4:06] um options around them, whether it was like, you know, a grocery store, restaurants, [4:12] or like cooking from home. And for the providers, what we do, we aggregate all the risk that we can [4:19] uncover in between visits. So, if you notice like your patient is not sleeping well, [4:25] or like, you know, their their heart rate it just keep jumping around, we aggregate that risk and we [4:29] flag it to the top and send it, you know, send it within like in a dashboard to your doctor or the [4:33] health care team. So, that way they can see in a real time, okay, there is something wrong going on [4:40] in between visits. So, we we don't have to wait until the next one to to work on it. We need to [4:45] call them like in real time. Uh when we start like in our journey like in a year ago, [4:53] uh LLMs were like, you know, super fancy and, you know, they were yeah, okay. So, they were so smart [4:58] that when you send them some prompt with their in a specific data, they output, you know, they're [5:03] good in making it it looks confident and like, you know, yes, this is what I'm looking for. So, [5:08] our previous model we started with was basically we could look at all the signals from different [5:14] inputs. Again, like, you know, smart hardware, the environment around you, where is your location, [5:20] uh your medication you're on, and like a different different single of input [5:24] or data points. We realized that then we wrap all this information within like in a single [5:30] single prompt with maybe 4,000 tokens or so, and we pass it to the LLM. And we say, okay, [5:36] what should be the recommendation? And, you know, it did work, but, you know, at one day, [5:41] uh we said, okay, now we uh we what what can we audit? Like, you know, the output, [5:47] how can we audit it? Can we provide the doctor and the patient with like what do you call [5:52] a verification receipt that, you know, can answer this question, why did you you say give me this [5:57] recommendation at this moment? And that's where like, you know, we find out now it is not working. [6:03] So, the prompt it may understand the question, but they cannot understand what state are you [6:08] going in right now as a patient. And that's, you know, this was the moment that we we decided to [6:16] sit down and ask, you okay, so find out what question were you asking the LLM to do for us. [6:22] And for the prompt, uh, you know, prompt based, uh, uh, uh, model, basically we were asking the [6:29] LLM to tell us what that what the person should do right uh, should do right now. [6:35] And that was the wrong question. So, basically we had to flip it upside down. We need to go [6:41] one step back and and define the state. So, this is where like, you know, the moment we said, [6:45] okay, we're moving from the prompt to what we call the state, uh, uh, uh, [6:51] AI model. So, the question we're asking right now, what is state is this person in right now? So, [6:57] this is the starting point for us. We need to understand what state are you going right now. [7:02] So, I'll give you an example. So, last night I did walk 20,000 steps. And but I did I did [7:08] my sleep at night was like, you know, breaking up. I had to wake up every two two hours. [7:13] If I have this information and send to the ChatGPT or one of the, you know, frontier LLMs, probably [7:18] they will come back and say, "Hey, congratulation, you did 20,000 step. You're an awesome. You're a [7:22] champion. Reward yourself with a big slice of cake." But the fact is like, you know, [7:27] my 20,000 steps, half of them was inside here. So, it was internal. And the other half was outside [7:35] in Vegas like a strip where the temperature was 113° and you know, I was walking around. So, when [7:43] when asking what the state this person right now, in fact, my state was like I was working [7:47] very hard last night. I was walking in like inside a building, and the other half was outside under [7:54] 113 degree. So, that's why my sleeping last night was not good cuz I was super fatigued. I like, you [8:01] know, it's not I'm trying to be active. Basically, I exhausted myself. This is why I couldn't sleep. [8:07] So, the recommendation that comes out of the runner right now, we're going to start you need [8:11] to hydrate. You need to drink more water. This is the first thing that you do in the morning. [8:15] Try to walk like less steps inside. Do not go outside. And we start recommending like nutrition [8:21] on how to recover, how to fix your joints because my joints is super fatigued right now. [8:26] So, this this is just example to tell you, like, you know, how how asking the right [8:30] question can make a huge difference, especially that we're fixing something for our health care. [8:36] It makes one one difference can make a huge trust issue for patients and doctors. [8:44] So, this is now like, you know, this is our our design right now. So, we have five steps. First of [8:49] all, we start with collecting the signals. We make sure that, you know, it's structured data. So, [8:54] we normalize everything. Then we apply like our rules and and and guidelines. [9:00] Then only then, once we have a specific definition of the state that our patient is going right now, [9:06] we can tell them our out of our recommendation tables, these are the qualified recommendation [9:12] for your state right now. Then we move we change the LLM job to be our communication agent. So, [9:19] it's not the reasoning anymore. It's only communication cuz this is what they are best in. [9:24] They can speak English, French, 100 language in the same in the same seconds, and they can make it [9:30] in a plain English. So, all the medical terms that we're using in between, we use LLM to simplify [9:36] it and send it like in a day-to-day message to that. Uh so, you know, the screen is small here. [9:45] Uh so, the benefits out of this now we can audit we can trace every single. So, [9:50] I can start from the LLM communication the problem with the output. I can ask my model, [9:54] "Tell me why did you recommend this, you know, this restaurant or food item from a restaurant [9:59] to my patient?" And it can tell me very easy because we have your the breadcrumbs like, [10:03] you know, all the way and from the input to the output that we can track back and and generate [10:09] this receipt for doctor and patient just to make make sure that, you know, they understand why [10:14] we recommend that for them and why we think this is the right uh for them. And uh we can [10:20] we can apply our own uh guardrails and rules, so it's not a black box anymore because LLM is [10:25] a black box like when a weak model doing too much. And also the personalization point of it. So, [10:32] we make sure that, you know, whatever is coming out of the model is only you, [10:36] not like a generic when you ask uh ChatGPT, "I'm on a GLP-1 medication, what should I eat?" [10:41] And they will tell you, "You need to eat more protein less carb." No, we do not do that anymore. [10:46] Uh another benefit like, you know, from the same state because we're thinking out of the [10:50] same model, we can help patients to get like in a personalized recommendation. And we have to [10:55] make sure that your doctor have visibility in this like in 90 days in between visits [10:59] so they are not falling into the blind spot and try to uncover like uh unhealthy stuff. [11:05] Uh at the end like, you know, this is the most like, you know, you know, advice I can [11:09] share with you like, you know, try to make your model understand first then speak second. So, uh [11:15] I know put your uh build the state for whatever customer you're working with retail, [11:21] patients, whatever like, you know, environment you're in understand what's your end user like, [11:25] you know, their state what their real need then apply your own boundaries then, you know, [11:29] let the LLM do what what they are best in doing which is basically you know, translate [11:34] this or communicate this information back to the patients. Uh, thanks so much. Uh, this is me, uh, [11:40] Yaya. Uh, we're at D-Runner. Uh, if you are on the the the conference app, you can just [11:46] look me up or you can send me an email at like support@d-runner.com. We'll be there. Thank you.