aoi-agent

aoi-agent is a re-verification layer for PCB automated optical inspection lines. Production AOI is tuned for recall and over-flags, so six in ten flagged regions are false alarms that a person reviews today. The project puts a vision model in front of that queue and an agent behind it, so only what neither can settle reaches an operator.

A ResNet-18 re-verifier scores each candidate region, and a LangGraph flow gathers context and a language model rationale for the uncertain ones, escalating what it cannot settle to an operator through a durable interrupt. A second entrance, /ask, turns a supervisor question into a validated plan of typed lookups with a chart. The stack is Python 3.12 with PyTorch, LangGraph, MCP, FastAPI and SQLite, and the default model runs locally through Ollama.

The README reports its own benchmark on the public DeepPCB set: 55.6 percent of manual review removed at a 0.66 percent escape rate, against a 0.5 percent budget in the project's own work instruction, so on that reading the budget is not met. The README is explicit that the demo runs on the DeepPCB set with a differencing simulator and seeded line records, not on a live production line, and that photographed boards fail the differencing front end at the first gate.