Watchdog

Inspiration

I volunteered at the zoo for three years, letting me learn a lot about animals and the illegal wildlife trade.

What it does

Watchdog reads through marketplace listings, using a dynamic agent system to decide whether the listing is likely to use illegal animal products. If so, the system checks listings with similar keywords and all other listings from the same seller.

How we built it

Python, FastAPI, Pydantic, Anthropic Claude API, Claude Sonnet 5.5, Claude Haiku 4.5, AGNTCY (SLIM, Agent Directory), A2A, Model Context Protocol, Next.js, React, TypeScript, Tailwind CSS, SQLite, Docker, eBay Browse API, Met Museum Open Access API, Pillow, imagehash, pytest, uv, Cloudflare Tunnel

Challenges we ran into

We struggled with testing our system and raising our accuracy.

Accomplishments that we're proud of

Our dynamic agent system, which calls agents when they are needed.

What we learned

We learned more about agents conversing with each other and reaching a consensus, which gave us a high accuracy.

What's next for Watchdog

Scaling the system for more marketplaces and other use cases, such as human trafficking.

Built With

• a2a
• agntcy
• anthropic
• claude
• claude-haiku
• claude-sonnet
• docker
• ebay-api
• fastapi
• imagehash
• met-museum-api
• model-context-protocol
• next.js
• pillow
• pydantic
• pytest
• python
• react
• server-sent-events
• slim
• sqlite
• tailwindcss
• typescript
• uv