Micius marks the defects and pulls the lot history; inspectors confirm before anything goes out

At a plant that makes custom parts for semiconductor equipment makers, a defect found at sampling often waits for a senior inspector to name it, and the lot’s history waits for a junior QA engineer to pull it from ERP and PLM, one system at a time. Zhiman gave both jobs to Micius, its model. Micius marks the defects for an inspector to confirm part by part, then pulls the records and drafts a report the inspector approves before it goes out.

A lot of custom parts comes off the line, and the sampled part has burrs along one edge and a small pit on its face. The young inspector can’t tell whether the pit came from handling or machining, so they take it to the senior inspector, who has been at the plant for years and can usually name the defect at a glance, and often the step it came from.

Then someone has to trace the lot: which supplier and material batch, which station and routing, which drawing revision and inspection spec. Those records sit in ERP and PLM, and a junior QA engineer digs them out one system at a time and writes them up while the engineers wait at the inspection station. Lots are often a few dozen pieces, with many part types and few samples of each, so there is rarely enough data to train a dedicated detection model. If the senior inspector is away, the uncertain parts sit and wait.

In its work with the plant, Zhiman had Micius take on both jobs: the first look at the photos and the search through the records. The inspector signs off twice, once on each part’s marking and again before the report goes to the engineers.

The inspector uploads the lot’s sampling photos and types “Analyse defects”. Micius matches each photo to a part in the lot and uses visual detection to look for burrs, scratches, dents, particles and porosity. Each part comes back as a card with the photo, the marked defects, a count by type and a one-line summary.

With so few samples of any one part, a photo alone can’t settle the call, so Micius waits for the inspector, part by part, to accept the marking, mark it as a false call, or send it for review by someone more experienced. Accepting confirms where the defect is and what it is. Pass/fail and lot release are decided separately, against the spec, by someone with that authority. In one anonymised lot of four sampled parts, the fourth had burrs along its right edge, two dents and a particle. Once the inspector accepted that marking, all four were confirmed and Micius moved on to the records.

Micius reads ERP and PLM for the material lot, supplier, machining station and routing, drawing revision, inspection spec, and the station’s calibration record. It only reads; nothing is written back. Each line can look fine on its own, so Micius checks whether the lines agree and notes which system each came from. Records that conflict, records that should exist but don’t, a station overdue for calibration: each is listed for a person to check. It doesn’t pick one version, and a blank never counts as a pass.

It also states the limits. A photo count matching the logged quantity doesn’t prove each photo is a different part. “Nothing found” covers only the defect types the detection looks for, not the back of the part or anything hidden. If a system returns only a spec number, Micius asks for the spec itself and leaves the limits blank.

In the same lot, two lines needed checking first: the material lot had no mill certificate in ERP, and the work order cited drawing rev B while PLM had released rev C. Both went into the report for the inspector.

Micius then drafts the report: defects and counts for each part, the lot’s history, the conclusion, what needs checking and what to do next. It can be edited and commented on in the browser and exported as a document. On causes, Micius only suggests where to look first and labels it a hypothesis; with burrs along one edge, it suggests checking the deburr step. It makes no scrap or release decisions and doesn’t touch inventory or work orders; a finished report doesn’t mean anything has changed in ERP.

When the report is ready, Micius lists the recipients, and nothing goes out until the inspector confirms. The approved report reaches the engineers in a messaging app, so they no longer wait at the inspection station. Reports the inspectors have reviewed can be kept as well. The next time a similar defect turns up, Micius can draw on those precedents, so part types with few samples slowly build up a record.

The plant also had a defect it couldn’t pin down: parts with the same routing and material still came out with porosity or scratches every so often. Finding it took the same skill as the inspection work, pulling scattered records together to see where they disagree. While chasing them, Micius called a weather lookup and brought each day’s temperature and humidity into the analysis. It pointed to the final drying step: its airflow needs adjusting with the day’s temperature and humidity.