FlowLab

The problem

Analysts who live in spreadsheets end up rebuilding the same load, clean, join, and summarize steps by hand, and then cannot explain where a number came from. FlowLab lets them compose those steps as a graph of nodes and run it on a server, so every result can be traced back through the graph.

The graph validates as you build it. A node that refers to a missing column turns red before anything runs.

The hardest stretch

Keeping two languages honest. Schema prediction lives in TypeScript so the canvas can flag problems at design time. Execution lives in Python and DuckDB. If the two ever disagree about a node's output schema, the UI lies to the user.

I wrote each concern once and made the disagreement detectable instead. At run time the worker asserts that the predicted schema equals the actual one for every node, and treats a mismatch as our bug, not the user's. In CI, a golden-fixture harness runs every node type through both sides and fails a pull request that has no fixture.

The AI repair loop taught me where this breaks down. The model cannot converge on a node whose warning only clears after a sample run, so pivot burned all four repair attempts on one unchanged diagnostic. I removed pivot from the AI-generatable catalog rather than hide the failure.

Decisions