Intelligent Medical Triage & Routing Agent for Low-Resource Clinics
# MedRoute
Intelligent Medical Triage and Routing Agent for Low-Resource Clinics
AMD Developer Hackathon ACT II -- Track 3 (Unicorn Track)
Voice-first, multilingual medical triage agent. Patients speak Urdu, Hindi, or English; Whisper Large V3 Turbo transcribes locally with native Urdu support and Roman Urdu normalization; a deterministic LangChain orchestrator reasons over symptoms with WHO-grounded RAG; model inference uses OpenRouter's free-model router for low-cost testing. A formal safety governance layer with 16 hard red-flag overrides runs before any LLM is invoked.
## Architecture
```
Voice Input --> Whisper Large V3 Turbo --> Input Parser (symptoms, clusters, duration)
|
Safety Pre-Check (16 hard red flags)
| (halt on match -> emergency)
Complexity Scorer
(syndrome clusters + calibrated confidence)
|
+----------------------------------------------+
| Deterministic Route Orchestrator |
| (routing policy is code, not free-form LLM) |
| |
| plan by route: |
| local_only -> OpenRouter |
| local_with_rag -> RAG -> OpenRouter |
| remote -> RAG -> OpenRouter |
| escalation_bias -> RAG -> OpenRouter |
| |
| cascade on failure: |
| provider down -> clinical heuristics -> |
| clinician escalation |
| |
| confidence fusion: scorer (union) model |
+----------------------------------------------+
|
PDF Report + urgency + cascade audit
```
**Design principles** (aligned with hybrid CDSS / CLARITY-style systems):
- LLMs support diagnosis; **routing and safety are deterministic**
- Complexity-aware grounding: simple cases call the model directly; complex cases add retrieved evidence
- Conservative under uncertainty (escalation bias + confidence fusion)
- Works offline: seed RAG + …