My family sells wine. It's a small shop where individual relationships are everything.

But attention gets spread thin when a few people are managing hundreds of customers' relationships with thousands of products. So I built an agent that knows the store. It pulls from our point-of-sale system into a database the agent can query, and answers plain-English questions about the business. Now, when an old friend reaches out, their entire history with the store is retrievable with one sentence of input.

Under the hood it's an orchestrator that decomposes a question into sub-tasks, fans the independent ones out across sub-agents concurrently, and streams back a synthesized answer with live-rendered charts. The model never touches the data directly: it reasons about what to ask, then works through typed, schema-validated tools. Everything is logged as an OpenTelemetry span into a local store.

The whole thing ships as a double-clickable macOS app with zero dependencies on the recipient's machine. The backend compiles to a single binary with Postgres embedded in-process (PGlite), and the app bundles the real dataset, topped up incrementally from the POS API on every launch. My parents open a window and ask it things.

I AI-fudged the image to obfuscate the private data! Hope you don't mind the fluff...