Shelter observations, routed to the people who need to know.

Shelter software has a field for everything that fits in a form and nothing for what staff actually notice. A volunteer sees a dog pee purple, or notices it was lethargic on a short walk, and there is nowhere structured to put that. It ends up in a group text, or nowhere. Even when it is written down, there is no clear rule for who should hear about it: sometimes the vet team, sometimes behavior, sometimes facilities. The shelter staff we talked to said there is no one clear-cut way to decide.

PetPilot is a shelter management front end with two ways in for those observations. A remarks router takes what a person noticed in plain English, classifies it, routes it to the right teams, and sends the alert. A Truth Queue takes what a camera flagged and holds it for a human to confirm before it becomes part of the dog's record.

Winner, AI Valley Dogathon 2026. Built in 5 hours by a team of 4.

Live version · Remarks router internals · DogOS spec

What it does

Where	What happens	Why it matters
Foster roster	Every dog with placement, status, location, and one click to health, location, remarks, and camera history	One place to see who is where and what is open
Remarks	Type an observation, hit Classify & route. Claude labels it, a routing table picks the teams, Arcade sends the Slack alert	The weird stuff finally lands with someone who can act on it
Truth Queue	Camera observations wait for a person to Confirm, Edit, or Pass	A camera guess never becomes a medical fact on its own
Medical	Month calendar of treatments, with overdue and due items marked	Rechecks and boosters stop slipping
Intake & outcomes	Month-to-date intake by source, outcomes, and recent movement	The numbers a coordinator gets asked for

How it works

Remarks: free text in, the right teams alerted

flowchart LR
  A["Volunteer types a remark<br/><i>peed purple this morning</i>"] --> B["Claude classifies<br/>one of six labels"]
  B --> C["Routing table<br/>label → teams"]
  C --> D{SEND_MODE}
  D -- slack --> E["Arcade sends<br/>Slack alert"]
  D -- log --> F["Logged, not sent"]
  E --> G["New row in the dog's Remarks<br/>label chip + who was told"]
  F --> G

Loading

Routing is a lookup table, not a model decision. That keeps it inspectable: you can read exactly why an alert went where it went, which matters more than sophistication when a vet is the one being paged.

Label	Routed to
medical_urgent	#shelter-medical, #shelter-managers
medical_routine	#shelter-medical
behavior	#behavior-team
facility	#operations
enrichment	#enrichment
unclear	#shelter-managers (when in doubt, a human triages)

Arcade handles the authenticated send, so the project ships no OAuth flow of its own and per-user tokens stay out of this codebase. A shelter alert router has to reach people across a volunteer base that turns over constantly, so delegated auth is the right primitive rather than an afterthought.

Truth Queue: a human confirms before anything becomes fact

flowchart LR
  A["Camera flags an observation<br/><i>missed feeding, pacing,<br/>location conflict</i>"] --> B["Truth Queue<br/>pending review"]
  B -- Pass --> X["Dismissed<br/>kept for audit"]
  B -- "Confirm / Edit" --> C["Deterministic triage<br/>route + priority"]
  C --> D["Case opened"]
  D --> E["Slack alert<br/>to the routed team"]
  D -- "shelter medicine" --> F["Calendar recheck<br/>in 24 hours"]

Loading

Triage is a small set of rules, not a model. It never produces a diagnosis, only a route (shelter medicine, behavior, operations, volunteer review) and a priority. A medical lean only applies when the dog has recent medical history, such as a feeding change right after a procedure.

A closer look

Medical calendar	Intake & outcomes

Running locally

You need Node.js and Python. For the Truth Queue you also need Docker (for MongoDB). Each service runs in its own terminal.

1. Front end

From the repo root:

python -m http.server 8000

Open http://localhost:8000/Foster%20Board.dc.html. The page must be served, not opened as a file; double-clicking the HTML leaves every {{ value }} unresolved because the runtime cannot boot.

2. Remarks router (port 3001)

cd remarks-router
cp .env.example .env              # Windows: copy .env.example .env
npm install
node --env-file=.env server.js    # prints "remarks router listening on :3001"

npm start does not read .env, so start it with --env-file as above.

Variable	Needed for
ANTHROPIC_API_KEY	Required. Without it, classification fails. Get one at console.anthropic.com
ARCADE_API_KEY / ARCADE_USER_ID	Only for real Slack sending
SEND_MODE	log classifies and logs the alert without sending. slack sends for real, once Arcade is set up and you have joined the channels

3. DogOS backend for the Truth Queue (port 4000)

cd backend
cp .env.example .env
docker compose up -d     # MongoDB on :27017
npm install
npm run seed             # 7 dogs, medical records, 3 pending observations
npm start                # prints "DogOS backend listening on :4000"

Without Arcade keys the backend runs in stub mode and logs the Slack and Calendar actions it would take. To point the page at a deployed backend, open it with ?api=https://your-backend/api; the choice is remembered in the browser.

.env files are git-ignored. Never commit them.

Windows: if npm fails with "running scripts is disabled," use Command Prompt instead of PowerShell, or run once: Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned

Try it

Open a dog's Remarks, type an observation, hit Classify & route. A new row appears with a label chip and the teams it routed to. If you see "Couldn't reach the router," the remarks router is not running or errored; a missing ANTHROPIC_API_KEY shows up in its terminal as an authentication error.

Then open Truth Queue and confirm one of the seeded observations to see the case it opens.

Where things live

Path	What
Foster Board.dc.html	The whole front end: roster, Truth Queue, medical calendar, intake and outcomes, and the record drawer. Single-file Design Component app, React via CDN
support.js	The Design Component runtime the page loads
_ds/	Design system bundle the page loads
remarks-router/	Classify, route, and send (Node/Express). See its README for endpoints and internals
backend/	DogOS backend (Node/Express + MongoDB): dogs, observations, review queue, cases, care events, Arcade actions
CLAUDE.md	The DogOS product and build spec
docs/assets/	Banner and screenshots for this README
.replit	Replit run config for the deployed backend

The front end calls the remarks router at http://localhost:3001 (REMARKS_API) and the DogOS backend at http://localhost:4000/api (API_BASE), both set near the bottom of Foster Board.dc.html.

Honest limits

Built in one 5-hour session, so:

• All shelter data is synthetic. There is no real PetPoint integration; PetPilot is designed as a layer that sits alongside a shelter's system of record, not a replacement for it.
• Remarks routing is a fixed table, not learned from a shelter's own history of who got alerted for what. That is the obvious next step, and the reason to want real data.
• Six labels is a starting taxonomy, not a validated one.
• Truth Queue triage is hand-written rules over six observation types. The camera observations are seeded; there is no live vision pipeline in this repo.
• The remarks router and the DogOS backend are separate services that do not share data yet. A typed remark does not enter the Truth Queue.

What this would unlock

The router has to touch every system a shelter already runs, which makes it a reasonable wedge for the things the shelter staff asked for and we could not build in five hours: syncing to electronic health records so an adopted animal arriving at a hospital does not need manual re-entry, keeping Salesforce and the shelter system consistent instead of diverging, inter-shelter transfer requests with tracking, and quarantine history that is currently hard to retrieve at all.

Team

• Pranav Balachander, pruhnav
• Kiran Oguri, kiran172
• Arshia Bhattacharyya, arshiabyya
• Trisha Moorkoth, tprofessional