The paper registers of every PHC already contain next week's outbreak. Nobody is reading them in time. We built the system that does — and put all 33 districts of Gujarat on one screen, ready to switch on.
A district-health command centre for Gujarat: AI intake at the facility, classical outbreak surveillance and medicine-redistribution analytics at the district, and a statewide coverage map. Dahod is the live pilot district; the other 32 are on the map, waiting to be switched on one block at a time.
Originally built for Build with AI: Code for Communities — Track 3: Smart Health.
Primary Health Centres and Community Health Centres run on paper. Daily OPD registers, stock cards and bed counts never leave the facility in usable form — the digital reporting that does exist (HMIS, IDSP/IHIP) is weekly or monthly, typed in late, and aggregated past the point of usefulness. Three consequences, all visible in a tribal district like Dahod:
- Outbreaks are detected late. When three centres in one block each see a small rise in fever-with-rash cases, that is a dengue signal days before any weekly report shows it. Today, nobody can see it.
- Medicines run out in one place while expiring in another. District officers discover stockouts when a centre calls, and expiries when the audit happens.
- The data burden falls on the least-staffed tier. Every new reporting requirement means another form for a pharmacist-cum-data-clerk who is also seeing patients.
State view (all 33 districts). Every district of Gujarat is on the map, drawn from the government's own health-centre directory — real names, real facility network. Dahod glows: real facility reports are flowing through the radar. The rest are dark, waiting. The pitch is the picture: light up Gujarat, one block at a time. We do not fabricate statewide activity — a real skeleton with one live district is more credible, and more honest, than 33 districts of invented numbers.
Dahod (the live pilot). The full command centre runs on Dahod's real facility network — District Hospital + SDH Devgadh Baria + 12 CHCs + PHCs across 9 tribal talukas (Dahod, Devgadh Baria, Dhanpur, Fatepura, Garbada, Jhalod, Limkheda, Singvad, Sanjeli).
Staff send the day's numbers as a voice note in Gujarati, Hindi or English or a photo of the paper register. Gemini's multimodal models parse either into a structured daily report (footfall, cases by IDSP-aligned symptom category, bed occupancy, stock on hand with expiries). The AI drafts, staff confirm — fields the model was unsure about are highlighted for human review before anything is saved. Every report is attributed to a named staff member with an on-the-spot photo.
Every confirmed report feeds four live analytics:
| Layer | What it answers | How |
|---|---|---|
| Outbreak radar | "What is flaring, where, since when?" | EARS-C2-family aberration detection (the statistics used by CDC/WHO surveillance): 21-day rolling baseline with a 7-day guard band per facility per syndrome, corroborated at block level — one noisy facility never raises an alert, three centres rising together do. |
| Stock emergencies | "Who runs out of what, when?" | Days-of-stock = on-hand ÷ 14-day burn rate, per drug per facility. |
| Expiry waste (₹) | "How much money is about to expire on shelves?" | FEFO simulation of every batch against its facility's burn rate; unconsumable units valued at procurement price. |
| Redistribution | "Which transfers fix both problems?" | Greedy matching: worst shortages first, nearest surplus holder, expiring stock offered before fresh stock, outbreak blocks pre-positioned to 21 days of cover for the relevant drugs. |
Plus a one-click weekly brief for the Collector/CDMO — Gemini turns the analytics into one page of plain language, in English, Hindi or Gujarati.
Arogya Radar is not a replacement for HMIS / IDSP / IHIP / e-Aushadhi. It is the last-mile layer that finally makes their daily reporting actually happen, then exports into them. It is read-only augmentation of what facilities already do on paper — it changes nothing they run, it only makes the register they already fill in visible in time to act. Adoption path: a no-cost, read-only pilot in one Dahod block, blessed by the CDMO/Collector, proving one outcome number (an outbreak flagged N days before HMIS; ₹X of expiry prevented; reporting compliance up).
A deliberate division of labour:
- Gemini does perception and language — reading messy registers, understanding code-switched Gujarati/Hindi voice notes, and writing briefs: the things nothing else can do.
- Detection and forecasting are classical, auditable statistics — a district health officer must be able to see exactly why an alert fired ("284 fever cases today vs a baseline near 94, flagged 3 days running, 5 of 7 centres in Jhalod block"). Real surveillance systems (CDC EARS, WHO EWARS) work this way, and so do we.
No GEMINI_API_KEY? The app runs in mock mode: intake and briefs return labelled canned outputs so the full flow always works; every analytic remains fully live.
The repo ships with a 90-day simulated history for Dahod's real facility network. Seeded into it:
- A dengue-like outbreak ramping in Jhalod block (tribal, near the MP border — real malaria/dengue territory) over the final 8 days — the radar first flags it 7 days before "today", days before it would appear in any weekly report.
- Near-expiry surplus — dengue NS1 kits at SDH Devgadh Baria worth ₹1.65 lakh, ACT courses at District Hospital Dahod, amoxicillin at CHC Dudhamali… ₹3+ lakh that current burn rates cannot consume.
- The radar's flagship recommendation writes itself: move the expiring dengue test kits from Devgadh Baria in the west to the outbreak in Jhalod in the east — expiry prevention and outbreak response in a single transfer.
- A reporting blind spot at PHC Chakaliya — silent 3 days, inside the alerting Jhalod block, the most dangerous kind.
- A one-day diarrhoea blip that the radar correctly does not escalate.
| Real | Synthetic |
|---|---|
| All 33 Gujarat districts; Dahod's 9 talukas and DH/SDH/CHC/PHC network structure | Daily facility-level activity (does not exist publicly — that gap is the product's reason to exist) |
| PHC/CHC names from the Gujarat Health Dept district- & taluka-wise list of 1,474 PHCs (15-08-2018) | Facility coordinates (approximate to taluka HQ pending a GPS import) |
| Statewide totals (33 districts, 1,474 PHCs, 273 CHCs, 23 SDH, 23 DH) from NHM / Gujarat Health Dept | The seeded outbreak, expiry and blind-spot scenarios |
| Medicine catalogue: NLEM 2022 subset, Jan Aushadhi-level prices; syndrome categories aligned to IDSP |
Sources: Gujarat Health Dept 1,474-PHC list · NHM Gujarat report · Dahod district health portal.
npm install
npm run generate # regenerate the 90-day synthetic Dahod history (deterministic)
cp .env.example .env # add GEMINI_API_KEY for real AI intake/briefs (optional)
npm run dev # http://localhost:3000npx tsx scripts/check.ts prints the full analytics pipeline (alerts, shortages, expiry, transfers) in the terminal — the fastest way to see the radar fire.
/— Gujarat state radar (33 districts, Dahod live)/district/dahod— the Dahod command centre/intake— the 30-second facility report (voice / register photo / manual)
gcloud run deploy arogya-radar --source . --region asia-south1 \
--allow-unauthenticated --set-env-vars GEMINI_API_KEY=<your-key>Four Google Cloud services doing real work:
- Gemini 2.5 (Pro + Flash, multimodal) — voice-note and register-photo parsing to strict JSON; multilingual brief generation
- Cloud Run — containerised, serverless deployment (Dockerfile included; the live public URL)
- Maps JavaScript API — the interactive district map: facility pins, outbreak-zone circles, and the case-intensity heatmap
- Directions API — real road route + distance/ETA for medicine transfers (falls back to OpenStreetMap/Leaflet if no Maps key is set)
Pilot roadmap: WhatsApp Business + Speech-to-Text intake channel (zero-install for PHC staff), Firestore for durable storage and auth, BigQuery for cross-district analytics, export to IHIP/IDSP/e-Aushadhi formats.
┌──────────────── STATE VIEW ────────────────┐
│ 33 Gujarat districts · Dahod live · rest │
│ dark, waiting to be switched on │
└───────────────────┬────────────────────────┘
│ drill into the live district
voice note (gu/hi/en) ─┐ ▼
├─► Gemini multimodal ─► confirm screen ─► daily record store
register photo ────────┘ (perception) (human) │
▼
┌──────────────┬────────────────┬──────────────┐
│ outbreak │ stock burn & │ redistribution│
│ radar (EARS) │ expiry ₹ (FEFO)│ (greedy match)│
└──────┬───────┴───────┬────────┴──────┬───────┘
▼ ▼ ▼
district map + alert feed + tables + Gemini weekly brief
data/gujarat.json— the 33-district statewide skeleton (real districts + centroids + Dahod live flag)data/district.json— Dahod's real facility network (talukas, CHC/PHC names from the govt list)data/generate.mjs— deterministic synthetic-activity simulator (calibration + seeded scenarios)src/lib/radar.ts— EARS-C2-style aberration detection with block corroborationsrc/lib/stock.ts— burn rates, FEFO expiry waste, transfer recommendationssrc/lib/gemini.ts— the only file that talks to an LLM (intake parsing + briefs, mock fallbacks)src/app/page.tsx— the Gujarat state radar;src/app/district/[id]— the district command centrescripts/check.ts— terminal sanity harness for the whole pipeline
Mann Sutaria — solo build.