Work

Ponno needed a way to turn messy buyer messages into catalog lookups and orders. I was the founding engineer on a team of three. We built retrieval on pgVector and ChromaDB, and tool-calling with the Vercel AI SDK so the agent could take the next step instead of only chatting. I left after the architecture and the first working path were in place.

What I did

• • Built a retrieval layer on pgVector and ChromaDB that maps messy intent onto product and order data
• • Made tool-calling reliable when schemas drifted, context ran out, or the model invented a parameter
• • Set code review and architecture habits for a three-person engineering team

Stack

What came of it

• • Left a working retrieval and tool-calling path, not a slide deck
• • Three-person team, with review and architecture in place