Five-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4 and 5 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 Research Organization Built Tray Inspection That Confirms Every Fault Across Camera Frames
Every tray that crosses the line, filmed, judged and followed from load to offload. The agents watch all four cameras, judge every tray instead of the sample a person had time to pull, confirm a fault only when it holds across frames, and bring a person the two decisions that stay theirs: what happens to a flagged tray, and where the line's own definition of a fault sits. Five stations, run as a loop: the threshold a person sets at station 5 is the floor station 2 judges against tomorrow.
What Was Broken
What We Built
Five stations and 10 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Line Capture
Nobody is asked anything hereAgents film every tray off all four line cameras, tie each one to its lot, and hold any footage too blurred or washed out to judge from. Nothing waits on a person here. A tray inside held footage is marked unjudged, never waved through.
Fault Finding
Nobody is asked anything hereAgents outline every tray, lid, label and well to the pixel, name each fault they find, and grade it against the severity scale the quality team wrote. Nothing waits on a person here. Every tray is judged, not a sample of them, and a mark below the confidence floor is held, never graded by guess.
Tray Tracking
Nobody is asked anything hereAgents follow each tray camera to camera from load to offload, confirm a fault only when it survives across frames, and compare the line's rhythm against its last thirty runs. Nothing waits on a person here. A mark seen in one frame is a reflection until it proves otherwise.
Hold Decisions
A person answers hereAgents match each confirmed fault against the alert rule book, route what a rule already covers to the line leads, and build the case for each tray a person has to decide: the clip, the lot history, and what the same fault cost before. This is where a person works. Everything before it exists so that three trays, not eighteen hundred, reach this screen.
Shift Review
A person answers hereAgents close the count on every tray the shift filmed, map where faults clustered by camera, stage and hour, and propose the threshold changes that would have caught what was caught late, each shown with what it would flag wrongly. A person accepts or rejects each change. Where the line's definition of a fault sits, and what the model learns from, is not something an agent moves 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:
Without a hold on dark or blurred footage, the system grades trays it cannot see. A fault on one of those trays passes as clean, and nobody knows the footage was unreadable.
Below the confidence floor the grader is guessing. If the guess still gets a severity score, nobody downstream can tell it from a real grade, and people act on it as if it were one.
When the tracker loses a tray and guesses which one reappeared, every fault after that lands on the wrong tray. A defective tray can ship under a clean tray's record.
If a score alone can release or hold a tray, one wrong threshold acts on every tray with no person in the path. A single bad number turns into a shift of bad decisions before anyone looks.
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.
line-capturevision/footage/<lot-id>/<tray-id>.mp4hold-bad-footagevision/footage/**, vision/held/**, the camera streams read-onlyfind-and-gradevision/findings/<tray-id>.jsonbelow-confidence-floorvision/findings/**, vision/footage/** read-onlytrack-and-confirmvision/tracks/<tray-id>.jsonfault-unconfirmedvision/tracks/**, vision/findings/** read-onlyalert-and-casevision/alerts/<alert-id>.json, vision/cases/<tray-id>.mdawait-your-callvision/alerts/**, vision/cases/**, everything else read-onlyclose-and-tunerules/proposed/<change-id>.json, reports/line-<date>.jsonaccept-threshold-changerules/proposed/**, reports/**, the shift history read-onlyStack
Every system this agent reads or writes.