one stray, start to finish

How it works

One StrayWatch AI with six agents in one room. The AI reads and compares; plain, tested code decides urgency and who goes. Every step is written to a case ledger anyone can replay.

The six steps

with a real example
  1. 1Spotted

    A citizen sends a photo + location (or types a landmark), or a road camera sees an animal lying in a lane.

    example: β€œkutta ghayal hai, pair se khoon”
  2. 2Read

    Triage agent: species, injured or sick, lying down, on the road. The reporter's words override a calm-looking photo; if the photo is unclear, it asks.

    target < 5 s Β· measured ~2–3 s
  3. 3Known animal?

    Identify agent compares registered animals within 1 km. Has a caretaker? They hear first.

    target < 3 s Β· measured ~1.2 s
  4. 4Routed

    Rules set urgency and pick the nearest right team.

    alert < 30 s
  5. 5Escalated

    No accept? Watch agent asks the next one, then the duty coordinator. The reporter is kept updated.

    3 min (HIGH) / 10 min (CHECK) per hop
  6. 6Closed + learned

    Closed only with a photo, checked by AI. Hotspot map and health diary updated.

    proof required

Urgency rules

code, not AI
LevelWhenWho is asked
HIGH Β· URGENTInjured, sick, trapped or lying animal on the road (it needs care, and it can cause an accident) Β· dead animal on the road Β· trapped animal anywhere Β· the person on the spot says it's hurt and it's on the roadCaretaker β†’ rescue NGO β†’ vet β†’ municipal cattle team (for cows) β†’ traffic police (busy road: guard the spot) β†’ volunteers β†’ duty coordinator. 3 minutes each. Driver signboard turns on.
NEEDS CAREInjured or sick animal off the road Β· dead animal off the road Β· animal lying in a lane whose condition is unclearSame order, 10 minutes each.
NO CASEHealthy animal standing, eating or wandering, on the road or notNobody is disturbed; it is only counted for the traffic hotspot map. The citizen can tap β€œIt's hurt” any time. StrayWatch never removes healthy strays.

Same input β†’ same urgency, every time: 13 unit tests in tests/test_rules.py.

One AI, six agents

why not one big model?
AI Β· vision LLM

Triage agent

Reads photo + text in 6 languages into a strict, validated record. Asks if the location is missing.

AI Β· vision LLM

Identify agent

Matches a photo to nearby registered animals and lost posters. Acts alone only at 95%+; otherwise asks a human.

AI Β· YOLOv8 at the edge

Camera agent

Flags an animal lying in a traffic lane and not moving: likely hurt, and an accident risk. Ignores healthy animals.

Rules + map

Dispatch agent

Applies the rules and distance, alerts the right responder, checks the proof photo.

Timers

Watch agent

Times every open case; escalates to the next responder, then the coordinator.

Data queries

Hotspot agent

Finds where animals keep getting hurt, and when: β€œadd lights and a speed breaker on Ratu Road; keep a rescuer on call at 6 PM”.

Why split it?
  • Safe: an LLM never sets urgency or picks the team; unit-tested code does.
  • Testable: each agent has its own test. Swappable: change a model without touching dispatch.
  • Accountable: every hand-off is in one ledger. See it live in the Agent Room β†’