Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4, 5 and 6 need a person. Step 4 is waiting for approval right now. Click it to open the full walkthrough.
Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.
How a Global Technology Company Built Defect Detection That Tells a Scratch From a Lens Smudge
Every device that comes down the line, from the moment it passes the cameras to the moment it is packed, reworked or scrapped. The agents photograph every device instead of one in ten, name what is wrong with the picture attached, tell a scratch on the device from a smudge on the lens, and hand a person only the close calls, the scrap signatures and the decision to stop a line. Six stations, run as a loop: the certainty floor a person sets at station 6 is the floor station 2 names defects against tomorrow.
What Was Broken
What We Built
Six stations and 12 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Line Capture
Nobody is asked anything hereAgents photograph every device from four angles as it passes the arch, re-shoot any frame the light spoiled while the device is still inside, and hold a lane whose camera has gone dark. Nothing waits on a person here. A device with no usable frames is marked unphotographed, never judged blind.
Defect Detection
Nobody is asked anything hereAgents read every frame of every device, find each listed part, name every mark with its size, its spot and the picture behind it, and mark a device unsure rather than force a call. Nothing waits on a person here. A call below the certainty floor travels as unsure, never rounded up to a defect or down to clean.
Finding Confirmation
Nobody is asked anything hereAgents follow each flagged device across all four cameras, confirm a mark that moves with the device, clear a mark that stays with a lens, and send anything seen once to a person undecided. Nothing waits on a person here. A finding with one frame behind it goes to the pull bench marked undecided.
Device Disposition
A person answers hereAgents release the devices confirmed clean, pull the confirmed defects to the rework bench with the ticket already written, and queue the close calls for the inspector, worst first. This is where a person works. Everything upstream exists so that six devices, not eighteen hundred, reach this bench.
Line Watch
A person answers hereAgents count every confirmed defect by station and hour against the line's own normal, catch the same defect repeating device after device, trace it upstream, and put the call to stop or slow in front of a person with the evidence attached. This is where a person decides. An alert can see the run; only a person can stop the line.
Shift Close
A person answers hereAgents check every device that entered the shift against where it ended, chart where the findings landed, propose the threshold and holding changes the shift's mistakes point at, and put each one in front of a person with its cost shown. A person accepts or rejects each change. The certainty floor decides what a person sees, so an agent never moves it quietly.
Where a Person Decides
Operating Model
This changes how work flows through the team.
What Transfers, What Must Be True
Failure Modes
What breaks this pattern:
When nobody refuses a frame too poor to judge, the device still gets a verdict. The log shows a clean pass built on pictures that show nothing.
Without a floor, an unsure call gets rounded up to a defect or down to clean. Good devices leave the line as scrap, and bad ones ship, under a record that says the system was sure.
A dirty lens puts the same mark on every device that passes it. When nobody connects the repeats, each mark counts as a real defect, and the line throws away good hardware all shift.
A device gets scrapped because a number crossed a line, and no person signed. Scrap is money off the line, and now nobody owns the loss.
Directional Outcomes
What the agent counts, and the station that counts it.
These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.
What Runs Where
Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.
frame-intakeframes/<serial>/<camera>.jsoncamera-darkframes/**, tickets/maintenance/**, the camera feeds read-onlydetect-classifyfindings/<serial>.jsonlisted-part-not-foundfindings/**, frames/** read-onlyconfirm-findingsconfirmed/<serial>.jsonone-frame-onlyconfirmed/**, tickets/maintenance/**, findings/** and frames/** read-onlypull-and-ticketdisposition/<serial>.json, tickets/rework/<serial>.mdawait-scrap-signaturedisposition/**, tickets/rework/**, everything else read-onlyline-watchalerts/<alert-id>.jsonline-stop-proposedalerts/**, confirmed/** and disposition/** read-onlyshift-closerules/proposed/<rule-id>.json, reports/shift-<date>.jsonaccept-threshold-changerules/proposed/**, reports/**, the shift record read-onlyStack
Every system this agent reads or writes.