It started out as a broken project I had to revive: the third-party API integration didn't work, and the model's responses came back as messy, unstructured text. I fixed the integration and built a structured request/response format so the recipe output parses reliably every time. While I was in there, I found a prompt injection hole too. I closed it with layered defenses on both the client and the server: I validate input before it's sent, keep what the user types separate from the system instructions, and let the model itself sort real ingredients from junk before it's allowed to cook anything up.

I split this one across two AI agents working at the same time. One did the whole frontend, the other did the whole backend. Before either touched code, I wrote out a shared TypeScript contract covering every WebSocket event, payload, and route, so neither side had to guess what the other was doing. Each agent tested its own half before calling it done, then I did one final pass myself to make sure both sides actually lined up. Under the hood, the server holds the source of truth for event order and resyncs a client automatically if it falls behind, and presence runs through Redis pub/sub so it can scale across more than one backend instance.

I extended it past plain search: results get enriched with contact emails and business descriptions, there's a CSV export endpoint, and an n8n workflow handles automated cold outreach on top, with lead deduplication, send throttling, and CAN-SPAM/GDPR-compliant opt-outs built in.