Shipped

Marunthagam

Tamil-First Offline ASHA Triage · Gemma 4 Good Hackathon

What It Is

A three-tier offline community-health intelligence system for Tamil-speaking ASHA workers in rural India. Tier 1 runs a 5GB Q4_K_M Gemma 4 E4B GGUF entirely on an Android phone — model, IMNCI protocol engine, AES-256 encrypted SQLite — with zero network dependency for a triage decision. Tier 2 is a PHC doctor console; Tier 3 is a district health dashboard with geohash heatmaps and cluster alerts.

Why It Matters

India's 940,000 ASHA workers triage 80M Tamil-speaking citizens with paper checklists, no connectivity, and 1:10,000 doctor ratios in many districts. Marunthagam is decision support — not replacement — that catches life-threatening cases before they slip through. The mandatory Tamil disclaimer "இது மருத்துவ ஆலோசனை அல்ல" is enforced at the JSON schema validation layer; a triage record missing it fails to write.

What I Built

Three KALAVAI-fused LoRA specialists (triage / dermatology / maternal) on Gemma 4 E4B with a lightweight MoE router that activates the right specialist per query
IMNCI protocol engine as deterministic safety floor below the LLM — only upgrades urgency, never downgrades; confidence < 0.70 auto-escalates one level
Multilingual safety classifier (Tamil / Hindi-Devanagari / Gujarati / English) with morphology-aware regex — fix lifted refusal rate from 78% to 100% on n=100 adversarial prompts
Clinical relabeling pass: hand-reviewed 113 triage GREEN cases against a clinician rater — 18% were under-triaged, all toward higher acuity. Relabel before retrain.
Three-tier dashboard: Android APK (Pixel 6 build) + Vite/React Tier 2 clinic console + Tier 3 district map with geohash heatmaps and 48h cluster alerts (≥3 YELLOW in 24h flag)
Published 3 specialist models + dataset + clinician label-quality CSVs on HuggingFace — byte-identical to local export, sha256-verified end-to-end

Tech Stack

Gemma 4 E4BUnsloth QLoRA (r=32)llama.cpp / GGUFAndroid (Kotlin)React / ViteAES-256 SQLitePython

Outcome

Held-out F1 = 0.6491 (n=131, seed 42), RED recall = 0.5833, missed-emergency rate 0/12 — no gold-RED case slips through as GREEN. Workstation TTFT 0.007–0.038s, 195–213 tok/s — two orders of magnitude past spec. Safety refusal 100/100. Submitted to Gemma 4 Good Hackathon. Apache 2.0.

What It Is

What I Built

Three KALAVAI-fused LoRA specialists (triage / dermatology / maternal) on Gemma 4 E4B with a lightweight MoE router that activates the right specialist per query

IMNCI protocol engine as deterministic safety floor below the LLM — only upgrades urgency, never downgrades; confidence < 0.70 auto-escalates one level

Multilingual safety classifier (Tamil / Hindi-Devanagari / Gujarati / English) with morphology-aware regex — fix lifted refusal rate from 78% to 100% on n=100 adversarial prompts

Clinical relabeling pass: hand-reviewed 113 triage GREEN cases against a clinician rater — 18% were under-triaged, all toward higher acuity. Relabel before retrain.

Three-tier dashboard: Android APK (Pixel 6 build) + Vite/React Tier 2 clinic console + Tier 3 district map with geohash heatmaps and 48h cluster alerts (≥3 YELLOW in 24h flag)

Published 3 specialist models + dataset + clinician label-quality CSVs on HuggingFace — byte-identical to local export, sha256-verified end-to-end