for partners and developers
Tech docs & status
What StrayWatch is built from, what works today, what doesn't yet, and what comes next. Source code: github.com/meenakshisatpathy20-del/straywatch. Demo video: watch it here.
Architecture
every input becomes one case; every agent works from one ledgerInputs
π± StrayWatch app (installable PWA)
π¬ WhatsApp add-on
π₯ Camera agent (YOLOv8 at the edge)
π©Ί Vet consults
Agent Room (FastAPI)
Channel adapters β one event format
π TriageπΎ Identifyπ£ Dispatchβ±οΈ Watchπ₯ Cameraπ Hotspot
βοΈ Urgency rules (unit-tested code)
ποΈ Case ledger Β· Postgres (SQLite locally)
π§ Model gateway Β· BharatCode API: qwen3.5-397b (vision), gemma-4-31b-it
Outputs
π¨ Responder alerts (in their language)
π Reporter tracking link
ποΈ City dashboard + hotspots
πͺ§ Signboard feed
What works today
honest status, Nov 2026| Feature | Status | Notes |
|---|---|---|
| Citizen reporting (photo, words or a typed landmark) | Working | No sign-up; GPS blocked β landmark confirmed on the map; offline queue; photos shrunk on the phone |
| 6 languages end to end | Working | English, Hindi, Bengali, Tamil, Telugu, Marathi: app labels, chat, rules (words in each script), tracking page, posters; pre-translated fixed messages |
| Triage + Identify agents | Working | Vision LLM, strict JSON validated field by field; caretaker first at 95%+ |
| Urgency rules: URGENT / NEEDS CARE / NO CASE | Working | Hurt on the road = URGENT; healthy animals are never dispatched; unit-tested |
| Right-team routing | Working | Caretaker β rescue NGO β vet β municipal team (cows, dead animals) β police on busy roads β volunteers (sample teams) |
| Watch agent + duty coordinator | Working | 3 min (HIGH) / 10 min (CHECK) per hop; reporter updated each hop |
| Photo-proof closing | Working | AI checks the proof photo shows the animal safe; reporter sees it |
| Case ledger, Agent Room, tracking link | Working | Every hand-off recorded and replayable |
| Camera agent β cases | Working on recorded video | YOLOv8n + ByteTrack: alerts only for an animal lying in a lane; healthy animals ignored |
| Hotspot agent + signboard feed | Working | Where animals keep getting hurt, by road and hour; driver warnings while an injured animal is on the road |
| Lost & Found | Working | Searches past and future sightings nearby |
| WhatsApp add-on | Built | Tested with simulated webhooks; waiting on Meta verification |
| Vet Care: online verified vets | Working | AI pre-check + reviewed care card, emergencies to 1962/clinic, live translated chat, video link, vet's note, follow-up; vet sign-up + verification |
| Rescuer tools | Working | Private reporterβrescuer chat, open cases near me, transport requests, animal hospitals near the spot; new helpers approved by a coordinator |
| Vet appointments | Working | Video or clinic slots for the week, no double-booking, vet notified in chat, booking page with video link, owner can cancel |
| Food & nutrition shop | Pilot | Server-priced catalogue, cart, pay on delivery, street-dog feeding packs to volunteer feeders; confirmed by phone (no online payment yet) |
| Phone notifications | Working | Web Push to responders and reporters with the app closed |
| Map view | Working | Cases, injury hotspots, cameras; locations rounded to ~1 km |
| SMS / IVR, Telegram | Planned | Same channel adapter |
| Payments for paid pet consults | Planned | Free during the pilot; UPI later |
| Redis event bus, PostGIS | Planned for scale | Today: in-process events, bounding box + haversine distance |
Measured
~2β3 sTriage agent reads a photo + message
11/11animals re-identified among look-alikes, ~1.2 s each (edited copies of photos)
~10 sreport β caretaker alerted, on the free live server
35 msCamera agent per frame on a 4-core CPU
0wrong-city or beyond-8 km alerts in a 1,238-report, two-city load test
78automated tests, run in GitHub Actions
Limitations
what we're careful not to claim- Responder teams on this site are samples. Real NGOs, vets and municipal teams join in the pilot.
- Camera AI uses a pretrained model: cows are reliable, street dogs are not yet (3/7). Fine-tuning on Indian road footage and night video is next.
- Re-identification was tested on edited copies of the same photos, which is easier than real life.
- The free server sleeps when idle; the first request can take up to a minute.
- The AI can misread a photo. That's why rules decide urgency, the reporter's words win, and people confirm matches.
What comes next
- 1Real partners
An NGO, the municipal team and 3β5 vets in two Ranchi wards; vet cards reviewed by a partner vet.
- 2Ranchi pilot
3 months, with hotspot cameras via the Smart City ICCC. Measure time to help and repeat injuries.
- 3More channels
WhatsApp live (Meta verification), then SMS / missed-call for feature phones.
- 4Scale
A new city is configuration: responders, vets, ward boundaries, languages.
API
| Endpoint | What it does |
|---|---|
POST /api/chat | One chat turn (text, button, photo or location). The app and website demo use it. |
POST /api/agents/camera/events | Edge Camera agent event (X-Agent-Key). Opens or merges a case. |
GET /api/room | Agent Room board: recent cases and agent activity. |
GET /api/signboards | Driver warnings for display boards near cameras. |
GET /api/map | Map layers (rounded to ~1 km). |
POST /vet/join Β· /admin/vets | Vet sign-up, and registration checks (admin key). |
GET /api/vet/slots Β· POST /vet/book | Free 30-minute slots for a vet and day; book one (a taken slot returns 409). |
POST /api/shop/order Β· /admin/orders | Place a food order (priced on the server); orders and bookings to confirm by phone (admin key). |
POST /webhook/whatsapp | WhatsApp Cloud API webhook (signature checked). |
GET /track/SW-n?k=β¦ | The reporter's private tracking page. |
pip install -r requirements.txt export ADAPTION_API_KEY=... # BharatCode model API uvicorn app.main:app --reload # http://localhost:8000 python -m pytest -q tests # offline tests with a stand-in AI