Testing the AI import on my own hardware

PayDown uses AI for exactly one job: reading pasted statement text and pulling out the plan details. Every number is computed by a deterministic engine — the AI never does math.

When it came time to test the parsing, the default path was the Claude API. I pushed back: I run Ollama on a box at home, and I wanted the extraction tested against my model, on my hardware, where iterating costs nothing.

Claude built me a small eval harness: sample statements (Klarna emails, Affirm statements, virtual-card schedules) run through my local qwen2.5 model, with the structured output diffed against known-correct values.

That harness turned out to be the best investment of the whole project. Over the following days it caught six-plus real extraction bugs from my actual statements — for free, before any user would have hit them. A $0 local model you can hammer beats a paid API you're afraid to iterate on.