The CoA agent is a web app for checking incoming raw materials. When a plant receives a lot, the supplier sends a Certificate of Analysis (CoA), a PDF of test results. You upload those PDFs and an agent reads each one and extracts the data. It then compares every result against the plant’s own, often stricter, specification and writes a report. The report rates the lot Conforming, Needs review or Non-conforming, with a reason for each test that doesn’t pass.
If the agent can’t tell whether a result passes, for example because of a unit mismatch, it pauses and asks you. It remembers your answer for future lots. The release decision always stays with a person: QA approves, holds or rejects the lot, and the decision is stored with the run.
The agent is built with LangGraph behind a FastAPI job queue, with Postgres and pgvector for storage and memory. The dashboard is Next.js, and its report page is generated by the model as a json-render spec. Sign-in uses WorkOS AuthKit. The plant limits in the repo are placeholder values, not real plant data.
