Procedure Drift Finder

Procedure Drift Finder helps plant safety managers see when real practice has drifted away from the written procedure, before it causes harm. Every factory already writes handover notes, near-miss reports and maintenance logs, but nobody has time to read hundreds of them against the standard operating procedures (SOPs). Our tool does that reading and shows the proof. The pipeline works in stages. Plain Python code splits each SOP into numbered steps and matches every shift note to the steps it might be about. An AI judge, running on [GPU] through an OpenAI-compatible vLLM endpoint with [MODEL], reads the note and proposes a drift type (skipped step, reordered steps, workaround or substitution, SOP outdated or impractical, SOP unclear) with the exact quote and a confidence score. Then plain code, not AI, checks that the quote appears in the note word for word. Any flag with an invented quote is dropped. A person then approves or rejects each flag. The Streamlit app shows a heat strip of which SOP steps drift most by week and shift, an evidence viewer with the highlighted quote beside the written step, a least-certain-first review queue, a benchmark with a list of our false alarms, and an upload page that turns Word, PDF, CSV and text files into the backend’s input format. We scored it against 25 planted drifts in 200 synthetic notes across 3 SOPs: 21% precision, 48% recall, 16 notes per minute. It reports on steps and shifts, never on people, and is not a safety control. All data is synthetic.