Stay

Stay is a free tool for dog owners who are close to giving up their dog over a behavior problem. An owner answers three short questions, and Stay explains what is likely driving the behavior and gives them one specific thing to try today. Thirty days later it checks back to see whether the dog is still home.

It is live at trystay.org and running in a pilot with Austin Pets Alive, with 50+ pilot users so far.

Why I built it

28% of dog surrenders are driven by behavior issues, and most of those are fixable with early, plain advice. By the time an owner walks into a shelter, it is usually too late for that advice to help. I wanted something that reaches them a few weeks earlier, for free, at the moment they start searching for help.

How a triage works

1. The owner describes the behavior across three short screens.
2. The backend sends the intake to Claude (Sonnet 5) with a fixed system prompt.
3. The response has to validate as a Pydantic model: a green, yellow, or red severity, the likely root cause, a first step, a plan for the week, and an escalation flag. A response that fails validation is logged and rejected instead of being shown to the owner.
4. The owner sees the result, can ask follow-up questions, and gets weekly check-ins plus a 30-day follow-up email.

Some cases should never get home advice first. If the intake mentions a bite that broke skin, a child involved, or an owner who is afraid of their dog, the result must be red, must be flagged for escalation, and must lead with a certified professional before anything else. That rule lives in the prompt and is checked by the eval harness below.

Engineering notes

• Tests and CI. 74 backend tests run on every push and pull request, alongside a frontend build. The suite exists because of a bug that sat in production for months: a template placeholder meant the owner's experience and prior training never reached the AI prompt. A regression test now fails if any unfilled placeholder reaches the model, for any intake shape.
• Evals. backend/evals holds a golden set of 10 cases, 3 of them safety-critical. Each case checks severity, escalation, and required referral tags. Scoring is pure Python and covered by the normal test suite, so CI verifies the harness without spending anything on API calls. Releases are gated on 100% recall of the bite and child safety rules.
• Prompt caching. The static system prompt sits before the cache boundary, which cut per-request token cost by 75%, confirmed with cache-hit telemetry.
• Rate limiting. The limiter keyed on the address it saw, which behind the hosting proxy was the proxy itself, so every owner shared one bucket and one busy hour would have locked everyone out. It now reads the real client address from the forwarded header, and the table of tracked addresses is capped so it cannot grow without bound.
• Security and privacy. Admin auth fails closed, dependencies are pinned, and a high-severity CVE in the router dependency was patched. Logging moved from print statements to structured logs, with owner data kept out of them.

Stack

Layer	Choice
Frontend	React and Vite, deployed on Vercel
Backend	FastAPI (Python), deployed on Railway
Model	Anthropic Claude API, Sonnet 5
Database	Supabase (PostgreSQL)
Email	Resend

Other features

• Dog profile with triage history and a behavior journal
• Shelter partner page with referral tracking, plus a weekly partner report (see docs/cron-weekly-report.md)
• Shareable result cards and behavior education cards
• Admin dashboard and an impact page

What we measure

The one number that matters is the share of owners whose dog is still home 30 days after their first triage.

Run it locally

Backend:

cd backend
pip install -r requirements.txt
cp .env.example .env   # set ANTHROPIC_API_KEY, SUPABASE_URL, SUPABASE_KEY
uvicorn main:app --reload

RESEND_API_KEY and ADMIN_PASSWORD are needed only for email and the admin dashboard. ALLOWED_ORIGINS (comma separated) replaces the CORS allowlist, which defaults to trystay.org, www.trystay.org, and the local Vite ports.

Frontend:

cd frontend
npm install
npm run dev

Tests and evals:

pip install -r requirements-dev.txt
pytest                                   # 74 tests, no API calls
cd backend && python -m evals.runner     # live evals, one API call per case
python -m evals.runner --safety          # safety-critical cases only

Author

Built by Harekas Bindra, CS and Linguistics at UT Austin. linkedin.com/in/harekas