Major Milestone Update: End-to-End Multi-Agent Ecosystem Deployed with Non-Negotiable Medical Guardrails and Live 4-Minute Interactive Simulation
We are pleased to share a major milestone in the evolution of Michelin: Veterinary Multi-Agent Clinical Ecosystem for feline internal medicine.
We have officially built, tested, and deployed an end-to-end, production-ready prototype featuring a dual-agent collaborative architecture, an atomic Central Patient Store (SSOT), strict clinical guardrails, and a live web simulation.
What Is New in This Release
1. Collaborative Multi-Agent Architecture
- "Michelin" (Veterinary Co-Pilot Agent — powered by Gemini 2.5 Pro): Collaborates directly with attending clinicians (Dr. Andy). Performs proactive clinical cross-examination, differential audits, and longitudinal lab trend analysis to prevent missed or delayed diagnoses.
- "Haiiro" (Caregiver Triage Companion — powered by Gemini 2.5 Flash): Empathetic, low-latency companion for pet parents at home. Ingests symptoms and photos, triages clinical risk (RED / YELLOW / GREEN), enforces non-negotiable emergency rules, and automatically escalates to Dr. Andy when thresholds are breached.
- Central Patient Store (SSOT): A unified, multimodal record with atomic version locking (record_version increment with conflict detection) that guarantees zero state drift between caregiver home logs and clinical charts.
2. Non-Negotiable Medical Guardrails
- Zero Hallucination Protocol: The system never interpolates or extrapolates numeric lab values or missing clinical data. Every cited metric is verified against the SSOT; unverified numbers are automatically redacted ([UNVERIFIED-VALUE-REDACTED]).
- Strict OCR Ambiguity Protocol: Any smudged, low-confidence, or blurred metric/image (such as Michelin's smudged platelet count "2?8") is forced to status: "AMBIGUOUS", value: null, with the mandatory clinical flag: "Data unreadable / ambiguous — manual clinical verification required".
- Deep Feline Domain Nuance:
- Longitudinal IRIS CKD Staging (2023): Catches creeping creatinine rising 1.4 -> 1.9 -> 2.3 mg/dL alongside inadequate urine specific gravity (1.022 < 1.035), staging the patient in IRIS Stage 2 CKD even when creatinine falls within standard laboratory reference intervals (0.6 - 2.4 mg/dL).
- Feline Triaditis Pattern Recognition: Correlates pancreatic elevation (Spec fPL 6.8 ug/L), hepatobiliary changes (ALT 148 U/L), and chronic gastrointestinal weight loss.
- Triage Downgrade Prevention: Deterministic feline emergency rules (breathing difficulty, male cat straining to urinate, 3+ vomiting episodes) can never be downgraded by the LLM.
3. Rigorous Test Suite
- 44 automated test cases passing with 100% coverage across type-level Pydantic models, strict OCR parsing, zero-hallucination citation verification, IRIS staging, multi-agent workflows, and FastAPI endpoints.
Live Links and How to Experience It
Experience the live ecosystem directly in your browser without installing anything or paying any cost:
| Resource | Link | Description |
|---|---|---|
| The 4-Minute Interactive Video Simulation | https://michelin-veterinary-ai.onrender.com/demo | Watch our 4-minute clinical walkthrough with synchronized split-screen views (Haiiro mobile triage + Michelin clinical co-pilot), live trend lines, and browser text-to-speech voiceover. |
| The Interactive API Explorer (Swagger UI) | https://michelin-veterinary-ai.onrender.com/docs | Test every endpoint live (triage observations, hypothesis cross-examination, and longitudinal trends). |
| Render | https://michelin-veterinary-ai.onrender.com | Production service running on Render Free tier with zero viewer cost. |
| Git hub | https://github.com/nadiajim/michelin-veterinary-ai | Full open-source codebase, test suite, and architectural documentation. |
Quick Tour of the 4-Minute Simulation (/demo)
If you visit the demo link, you will see a 5-act journey through the real patient case of Michelin (10.5yo male DSH):
- Act 1 (0:00 - 0:45): The subtle creep of feline chronic kidney disease masked by standard lab reference limits.
- Act 2 (0:45 - 1:45): Haiiro in action at home, handling a low-light blurry vomit photo with strict ambiguity guardrails and triggering a YELLOW emergency escalation.
- Act 3 (1:45 - 2:30): Central Patient Store (SSOT) synchronizing data atomically (v7 -> v8).
- Act 4 (2:30 - 3:30): Michelin Co-Pilot challenging Dr. Andy's gastritis hypothesis and flagging concurrent Triaditis.
- Act 5 (3:30 - 4:00): Medical guardrail telemetry and zero hallucination guarantee.
We welcome feedback and invite you to explore the live simulation and API endpoints.
Log in or sign up for Devpost to join the conversation.