SomaPulse 🩺

Offline-first medical triage and diagnostic translation at the edge — zero cloud, zero network, zero excuses

When disasters strike, communication grids collapse first. Rural health workers are stranded without access to cloud AI tools, unable to match colloquial symptom descriptions ("running stomach", "fire in the chest") to formal clinical protocols ("gastroenteritis", "GERD").

SomaPulse brings clinical intelligence to the edge. It transcribes spoken symptoms using local Whisper.cpp, maps colloquial terms to medical ontologies via SapBERT embeddings, and retrieves matching protocols from a local sqlite-vec database — all in under 120ms with zero internet.

Key Features:

• 🎤 Edge Speech Ingestion: HTML5 audio → local Whisper.cpp transcription (38ms p50)
• 🧬 SapBERT Semantic Mapping: Aligns colloquial dialect terms to UMLS clinical concepts
• 💊 Protocol Retrieval: Cosine similarity search against pre-seeded medical guidelines
• ⚠️ Drug Contraindication Warnings: Critical interaction alerts rendered instantly
• 🔌 100% Offline: No network, no cloud API, no subscriptions — runs on a $300 laptop

Click to expand all dashboard screenshots

Idle State — Ready to Triage

Edge system fully operational: Whisper 16-bit GGML loaded, SapBERT ONNX v1.2 ready, 142 protocols indexed, 1180 MB RAM on i5-8250U. Voice waveform visualizer idle. Quick-access demo queries for Diarrhea, Fever, and Breathing visible below the recorder.

---

Diarrhea Triage — "running stomach"

Transcript: "having severe watery stool and cramping for two days". Matched Acute Diarrhea & Dehydration Protocol (Gastroenteritis) at 91.2% similarity in 118ms. Action steps include ORS administration and zinc dosage. ⚠️ Contraindications: Do NOT administer antimotility drugs if bloody stool observed.

---

Breathing Triage — "chest whistle"

Transcript: "tight breathing heavy lung chest whistle sound for three hours". Matched Acute Bronchial Constriction Protocol (Asthma/Bronchospasm) at 85.6% similarity in 102ms. Action steps: sit upright, salbutamol inhaler, monitor respiratory rate. ⚠️ Do NOT use sedatives.

---

Fever Triage — "hot body / skin fire"

Transcript: "my child has a hot body shaking skin and cannot swallow water we are at a high elevation camp". Matched Severe Hyperpyrexia & Acute Tonsillitis/Meningitis Protocol at 88.7% similarity in 94ms. 5 action steps including HIGH ELEVATION WARNING. ⚠️ Three contraindications including meningitis rule-out and immediate transport.

graph TD
    Responder([Field Responder]) <-->|Audio Recording / Web UI| UI[Next.js 16 / React 19 Frontend]
    UI <-->|Local API calls| API[Python FastAPI Backend]
    API <-->|Speech Stream| Whisper[Whisper.cpp <br/> GGML 16-bit Quantized]
    API <-->|Extract Symptoms| ONNX[SapBERT ONNX Runtime <br/> 384-dim Biomedical Embeddings]
    API <-->|Vector Distance Query| SQLite[SQLite + sqlite-vec <br/> Cosine Similarity Search]

Data is persisted in Supabase (PostgreSQL) with Row-Level Security enabled. All tables use the sp_ prefix to namespace within the shared Supabase instance.

erDiagram
    sp_triage_results ||--o{ sp_protocol_matches : "produces"
    sp_triage_results {
        serial id PK
        int latency_ms
        text transcript
        timestamptz created_at
    }
    sp_protocol_matches {
        serial id PK
        int triage_result_id FK
        int protocol_id
        text title
        text clinical_term
        numeric similarity
        jsonb steps
        jsonb warnings
    }
    sp_system_health {
        serial id PK
        varchar status
        varchar network
        boolean whisper_loaded
        boolean sapbert_loaded
        int db_records
        int ram_usage_mb
        text cpu_model
    }
    sp_vocabulary_gaps {
        serial id PK
        text colloquial
        text clinical
        text umls
    }

RLS Policy: Anonymous read access enabled on all tables. Write operations require service_role key.

• Node.js ≥ 20
• npm
• Python 3.12 (for backend + benchmark scripts)

git clone https://github.com/edycutjong/somapulse.git
cd somapulse
npm install
cp .env.example .env.local
npm run dev

python seed.py   # Creates protocols.db with 142 pre-compiled SapBERT protocol vectors

python bench.py  # Validates full pipeline meets <200ms target (p50: ~58ms, p95: ~71ms)

48 passing tests across 4 test suites — covering mock data integrity, component rendering, interactive triage selection, voice recording state, vocabulary mapping validation, UMLS code format checks, and edge/offline system health branching.

npm test              # Run all 48 tests
npm run test:coverage # Coverage report
npm run lint          # ESLint
npm run typecheck     # TypeScript check
npm run build         # Production build
npm run ci            # Full CI pipeline (lint + typecheck + test + build)

CI runs on Node.js 20, 22, and 24 via GitHub Actions on every push.

devpost-uoe-somapulse/
├── docs/              # README assets
├── src/
│   ├── app/           # Next.js pages + __tests__/
│   └── lib/           # Mock data & utilities + __tests__/
├── .github/           # CI workflows
├── bench.py           # Pipeline latency benchmark script
├── seed.py            # Deterministic protocol database seeder
├── .env.example       # Environment template
├── LICENSE            # MIT
└── README.md          # You are here

Pre-Seeded Knowledge: The model runs entirely offline. It cannot answer open-ended medical inquiries outside the pre-seeded protocol indexes. Novel conditions require a new offline sync to expand the local database.

• Next.js 16 — App Router, React Server Components
• React 19 — UI framework
• TypeScript — Type-safe JavaScript
• Tailwind CSS v4 — Utility-first styling
• Python 3.12 — Backend runtime
• FastAPI — Async REST API server
• Whisper.cpp — Offline speech-to-text (GGML)
• SapBERT — Biomedical entity embeddings (ONNX)
• sqlite-vec — Vector similarity search
• SQLite — Embedded local database
• Jest — Testing framework (48 passing tests)
• GitHub Actions — CI/CD pipeline
• Vercel — Frontend deployment

MIT © 2026 Edy Cu

Built for UOE Summer of Code 2026. Thank you to the organizers and judges for the opportunity.