01 / Gemma 4 Hackathon · 2026
ImciFlow — grounded pediatric triage
LLM extraction, RAG evidence and deterministic IMCI rules in one auditable workflow.
ImciFlow explores a safer architecture for high-stakes AI: the model structures multilingual intake, retrieval surfaces source evidence, Python tools keep safety-critical classification deterministic, and the interface exposes missing information instead of hiding uncertainty.
Evidence register
3
languages
English, French and Sudanese Arabic paths
15/15
rule cases
Deterministic IMCI evaluation, not a clinical trial
2
live services
Vercel frontend and Cloud Run API
01 / Problem
Clinicians in constrained settings need fast protocol lookup, multilingual intake and explicit safety checks without delegating the final decision to a generative model.
02 / Approach
A graph routes intake through structured extraction, local retrieval, deterministic IMCI tools and a verification step before producing an explanation for human review.
03 / Outcome
The full-stack prototype is deployed and demonstrates traceable evidence, streaming pipeline events, audio/video inputs and an online/offline model-routing concept.
System trace
How the evidence is produced.
Designed and implemented as an end-to-end hackathon system covering product framing, orchestration, backend, frontend integration, evaluation assets and cloud deployment.
- 01Multimodal intake → structured clinical signals
- 02IMCI retrieval → cited protocol evidence
- 03Deterministic safety rules → triage class
- 04Human review → final decision
Validation scope
Backend, frontend and safety tests are present. A 15-case fixture evaluation verifies the deterministic classification layer after symptoms have already been structured.
Known limitation
The evaluation does not yet measure end-to-end extraction quality, multilingual robustness or clinical outcomes. The hosted demo is research software and must not be used for diagnosis.
What is inspectable
- FastAPI backend and React/Vite interface deployed separately.
- Safety-critical rules are kept outside the LLM path.
- Sessions capture model route, evidence, timing and safety flags for auditability.
Next proof to add
- 01Move verification after translation and expand citation checks.
- 02Add authentication, rate limiting, encrypted durable storage and retention controls.
- 03Benchmark extraction, retrieval, multilingual fidelity, latency and cost end to end.
Main stack