AI-Powered Virtual Village Clinic
A trained health worker at a village sub-centre. No resident doctor. This
platform closes that gap.
Core product principle: AI prepares the case. The doctor makes the medical
decision.
Built for Smart India Hackathon 2026 — Problem Statement 3: AI-Powered Virtual
Clinic for Rural Healthcare.
What it does
A health worker registers a patient by Aadhaar, records symptoms by voice in
Hindi or another Indian language, enters vitals, photographs the patient's paper
prescription and any visible injury, and runs an assessment.
A deterministic rule engine tiers the case. A classifier trained on 244,938
labelled symptom vectors proposes ranked disease candidates. A language model
writes the doctor-ready summary — and is forbidden from naming a diagnosis or a
medicine. The case is handed to a named doctor in the same district, who reviews
it, consults over video if needed, and signs the prescription.
High-risk cases skip all of that and go straight to the nearest of 75 real
district hospitals.
Why it is different
Most entries to a problem like this wire a public LLM to a prompt. This one has a
trained model, deterministic clinical rules that override it, and safety
boundaries enforced in code rather than asserted in a slide.
| |
|---|
| A model with a measured accuracy | Bernoulli Naive Bayes over 244,938 labelled symptom vectors — 377 symptoms, 582 diseases. Top-1 0.854, top-5 0.974 on 48,988 held-out cases. The artifact is committed to this repository |
| The LLM is bounded by the trained model | It may re-rank and reject the classifier's candidates. It may not introduce a disease outside them. Output stays traceable to training data, not to the model's imagination |
| Rules override the model, always | final_tier = MAX(rule, vision, model). A model can raise a tier and can never lower one |
| Medication is withheld from the health worker entirely | The problem statement permits protocol-based OTC suggestions. This system builds that engine, gates it eight ways, tests it — and then does not show it to the operator. A health worker acting on a drug name from an automated summary is the specific outcome this prevents |
| Nothing is ever fabricated | Silence returns no transcript, not a plausible sentence. An unreadable document returns no extraction. No hospital bed count is displayed, because no real-time feed exists |
| It says what it does not know | A flat posterior over 582 classes reports confident: false rather than five near-tied guesses presented as a shortlist |
| Honest about what is not validated | The triage thresholds and the formulary are drawn from published guidance and have not been reviewed by a registered practitioner for this deployment. The system says so on its own pages |
Key features
Multilingual speech capture — Hindi, English, Tamil, Telugu, Marathi,
Bengali, Gujarati, auto-detected. Three independent gates reject Whisper's
hallucinations on silence, because a fabricated symptom is worse than a missing
one.
Document OCR with mandatory human verification — prescriptions and multi-page
lab reports, read natively including PDFs, with all pages in one model call so
cross-page context survives. An extraction is a draft until a human confirms
it.
Health-card reading — an ABHA or health card photograph pre-fills
registration as a proposal the operator accepts field by field. Stores nothing.
A name of pure digits, a future date of birth or an unknown gender is refused,
and the API reports which fields it refused.
Wound-photo observation — observational only, never diagnostic, confidence
capped at "moderate" in code as well as in the prompt. Photographs live in a
private bucket behind one-hour signed URLs.
Deterministic clinical triage — seven age bands from PALS and WHO IMNCI,
automatic Celsius conversion, IMNCI dehydration assessed as a syndrome, and the
rule that a reading of 0 is a reading rather than an absence.
Doctor queue and review — worst-first, day-wise, scoped to cases assigned to
that doctor. A review requires a diagnosis. The decision travels back to the
health worker in real time.
Video consultation — peer-to-peer WebRTC over the platform's own
authenticated WebSocket. No third-party SDK, no per-minute cost, and consultation
media never traverses a vendor.
Emergency referral — 75 UP district hospitals by real coordinates, nearest
computed locally, no paid mapping API. Works when the link is bad, which is when
it is needed.
Six-role administration — super, state and district admins, doctors, health
workers and a read-only auditor, each scoped by a database column rather than a
UI filter. Administrators cannot read a patient record at all — enforced at
the router and again in row-level security.
Architecture
flowchart TB
subgraph B["Browser — React 18 + Vite"]
UI["Health worker · Doctor · Admin"]
end
subgraph C["One container"]
API["Express API — 64 routes<br/>+ /realtime WebSocket"]
PY["Python inference service<br/>FastAPI on 127.0.0.1:8001"]
API <-->|loopback| PY
end
subgraph D["Supabase"]
PG[("Postgres — 18 tables<br/>~40 RLS policies")]
AUTH["Auth"]
ST["Storage — private"]
end
subgraph M["Hosted models"]
G["Groq · gpt-oss-120b<br/>whisper-large-v3-turbo"]
GM["Gemini · gemini-3.6-flash<br/>+ fallback chain"]
QD[("Qdrant · protocols")]
end
UI <-->|HTTPS + WSS| API
API --> PG
API --> AUTH
API --> ST
API --> G
API --> GM
API --> QD
The assessment pipeline, in order — the order is the safety argument:
deterministic rules → vision severity → trained classifier
↓ ↓ ↓
sets the floor may raise only bounded candidates
↓
LLM synthesis — may raise, never lower;
may re-rank candidates, never invent one;
may not name a medicine
↓
medication deleted from every source
↓
doctor-ready handoff + tier workflow
Tech stack
| Layer | Choice |
|---|
| Backend | Node.js 22, Express 4, ESM |
| Database | PostgreSQL via Supabase — 18 tables, ~40 RLS policies |
| Auth | Supabase Auth for passwords, own 12-hour JWT for sessions |
| Frontend | React 18, Vite 5, Tailwind 3, Three.js (landing page only) |
| Realtime | One authenticated WebSocket carrying notifications and call signalling |
| Video | Peer-to-peer WebRTC with perfect negotiation |
| Inference | Python 3.11, FastAPI, scikit-learn |
| Models | Groq openai/gpt-oss-120b, whisper-large-v3-turbo; Google gemini-3.6-flash |
| Retrieval | Qdrant Cloud, filtered on approved = true |
| OCR | Gemini multimodal, with a Tesseract.js fallback |
| PDF | pdfkit, server-side |
| Tests | Jest — 135 tests, 9 suites — plus pytest for the symptom matcher |
| Deploy | Railway + Nixpacks, one container |
Every choice, and the alternative it was made over, is in
Tech Stack.
Quick start
Prerequisites: Node.js ≥ 22, Python 3.11, a Supabase project, a Groq key, a
Gemini key.
git clone https://github.com/PriyamMishra853/RURALAI-.git
cd RURALAI-
cp backend/.env.example backend/.env
cp frontend/.env.example frontend/.env
Fill both in — every variable is documented with its purpose in
Setup Guide §4.
cd backend && npm install
cd ../frontend && npm install
cd .. && bash ./build-ai.sh
Apply the database, in order:
cd backend
npm run inspect # what is in the target now
npm run db:apply -- --confirm # DESTRUCTIVE — runs 01–06
⚠️ db:apply currently stops at migration 06. Apply
database/v2/07_case_handoff.sql through 10_admin_analytics.sql by hand —
see Setup Guide §6.
Without them, case handoff, case withdrawal, emergency registration and the
admin dashboard are all broken.
Seed and run:
npm run seed # 36 states, 75 UP districts, 375 doctors, 1,875 patients
npm run seed:schedules # without these, every booking date reads "Closed"
npm run seed:daily # 5 cases per doctor, deterministic per date
[email protected] ROOT_ADMIN_PASSWORD='a long unique password' npm run seed:root
cd ../frontend && npm run build
cd .. && bash ./start.sh # http://localhost:5000
Verify:
cd backend
npm test # 135 tests. No network, no AI quota spent
npm run check # every external service reachable, models still exist
npm run preflight # is the system demo-ready right now?
Full instructions, including production deployment:
Setup Guide.
Documentation
For SIH 2026 evaluators: PS 26133 coverage and handoff ·
Pitch guide · Startup plan ·
Setup guide · Architecture
Recognition: 1st prize, IBM BOB Hackathon (CSJMU Kanpur) · 1st prize, Ideathon (HBTU Kanpur) ·
presented at AI MANTHAN 2.0 to the Governor of Uttar Pradesh.
Complete technical documentation is in
PROJECT_DOCUMENTATION/ — written from
the code, with every claim naming the file, table, endpoint or script that backs
it.
| |
|---|
| Architecture and Workflow | The whole system as diagrams — deployment, roles, clinical workflow, AI triage chain, consultation lifecycle, referral path, data model, API surface |
| 00 — Project Overview | Every built feature, the problem-statement coverage table, feasibility, scalability, sustainability, government-integration path |
| 01 — Setup Guide | Clone to production, followable from zero |
| 02 — Dependencies | Every package, with the reason it is there and what breaks without it |
| 03 — Data Flow | Speech, OCR, image, triage and realtime paths, traced separately |
| 04 — Workflow | The clinical journey, and each role's own sequence |
| 05 — Directory Structure | Annotated tree, every significant file in one line |
| 06 — Database Schema | Every table, column, constraint, enum and index; ER diagram; migration history |
| 07 — Tech Stack |
Scale demonstrated
| |
|---|
| States and union territories | 36 |
| Districts (Uttar Pradesh) | 75 |
| District hospitals with real coordinates | 75 |
| Doctors | 375 |
| Health workers | 75 |
| Patients | 1,875 |
| Visits, risk-tiered | 1,875 |
| Disease classes | 582 |
| Symptom vocabulary | 377 |
| Clinical alias mappings | 280 |
| Labelled training vectors | 244,938 |
| API routes | 59 |
| Database tables | 17 |
| Automated tests | 135 |
Status and honesty
This is a working system, deployed and answering. It is also a demonstration
system, and the documentation is explicit about the difference.
Not yet validated: the triage thresholds and the five-entry formulary are
drawn from WHO EML, NLEM, NEWS2, PALS and WHO IMNCI, and have not been
reviewed by a registered medical practitioner for this deployment. Medication
output is therefore suppressed by default in production, and the platform carries
a "not for clinical use" notice on its public pages.
Known gaps — including defects found while writing the documentation — are
listed in
Known Limitations and Risks,
with severity and what would close each one.
Section 12 is a roadmap. No fine-tuned model exists, no GPU has been
provisioned, and no clinical evaluation has been run. It says so in its own
opening lines.
Safety and legal notice
AI assistance does not replace professional medical diagnosis or treatment. All
clinical decisions, prescriptions and referrals are made by qualified
healthcare professionals registered under the National Medical Commission of
India.
This is a demonstration system. Its triage thresholds and medication list are
drawn from published guidance but have not been reviewed or approved by a
registered medical practitioner for this deployment.
Credits
Priyam Mishra — Core Developer
License
MIT.
