AI Inference Control Plane
A local-first engineering toolkit for measuring AI inference cost, comparing model configurations, and evaluating the cost of accepted work.
The engineering question
A cheaper model call is not always cheaper successful work. This project connects token usage, retry costs, and acceptance outcomes so engineers can compare configurations against the same tasks.
Inside the toolkit
The Node.js source includes exact token pricing with BigInt arithmetic, reviewed pricing snapshots, attempt and outcome tracking, bounded benchmark execution, and paired statistical comparisons. A local decision dashboard makes costs, missing evidence, and model-selection tradeoffs visible.
Current scope
This is a single-process engineering toolkit with public source and a synthetic demonstration. The demo uses illustrative models, prices, and outcomes. Routing views support planning; they do not automatically change production traffic.