Home Use Cases Work Insights About Contact
Back to Our Work

How a Global Research Organization Built Tray Inspection That Confirms Every Fault Across Camera Frames

Operations2
Open the walkthrough

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.

The Transformation

How a Global Research Organization Built Tray Inspection That Confirms Every Fault Across Camera Frames

Before
A person used to walk the line and eyeball trays in passing.
A person used to pull trays off the line and check them by hand.
A person used to re-check every flag the old spot checks raised.
After
every tray filmed, tied to its lot, blur held
every tray judged, every fault outlined and graded
a fault confirmed across frames, or rejected as a reflection
three trays waiting on your call, the clip on screen
the shift closed, two threshold changes waiting on you
Executive Summary

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

A person used to walk the line and eyeball trays in passing
A person used to pull trays off the line and check them by hand
The real cost
A person used to re-check every flag the old spot checks raised.

What We Built

Five stations and 10 subagents. Each subagent carries its own tasks and its own refusal.

A4
Frame Capture
Capture every tray passing the four line cameras, at line speed
A3
Footage Check
Check every clip for blur, glare, and a camera knocked off its mount
A4
Object Finder
Find every tray, lid, label and well in each frame, and outline each one to the pixel
A3
Fault Grader
Classify each fault it finds: cracked well, lifted lid, spilled sample, missing label
A4
Tray Tracker
Follow each tray across all four cameras from load to offload
A3
Pattern Watch
Check each found fault across every frame it appears in
A3
Alert Desk
Match each confirmed fault against the alert rules the quality team set
A1
Case Builder
Draft the evidence for each waiting tray: the clip, the frames, the lot and its history
A4
Shift Close
Count every tray the shift filmed, judged, flagged and decided, to a clean end
A2
Threshold Proposal
Propose the threshold or rule change that would have caught each repeated miss

How It Runs

1

Line Capture

Nobody is asked anything here

Agents 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.

2

Fault Finding

Nobody is asked anything here

Agents 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.

3

Tray Tracking

Nobody is asked anything here

Agents 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.

4

Hold Decisions

A person answers here

Agents 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.

5

Shift Review

A person answers here

Agents 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

Step 4, Hold Decisions. Decide hold or release on the 3 trays waiting. Say whether the film across the wells needs a rule of its own. a tray's fate is a person's call.
Step 5, Shift Review. Accept or reject the two proposed changes. Sign the 12-clip retraining set, or strike the clips that should not teach the model. the thresholds are yours to set.

Operating Model

This changes how work flows through the team.

Role
Responsibility
Hold Decisions owner
Decide hold or release on the 3 trays waiting. Say whether the film across the wells needs a rule of its own. a tray's fate is a person's call.
Shift Review owner
Accept or reject the two proposed changes. Sign the 12-clip retraining set, or strike the clips that should not teach the model. the thresholds are yours to set.
Line Capture, when it goes wrong
Look here only if a camera view looks wrong on the wall screen.
Fault Finding, when it goes wrong
Look here only if a fault reaches offload unfound.
Tray Tracking, when it goes wrong
Look here only if the line feels wrong and no alert has fired.

What Transfers, What Must Be True

What transfers
A person is in the loop wherever a tray's fate is decided or the line's definition of a fault moves. Two stations refuse to proceed on their own, Hold Decisions and Shift Review, and those are the two a person carries. Everywhere else the agents film, judge, track and route, and reach a person only when the call is theirs to make.
Hold Decisions stops for a person, and a tray's fate is a person's call.
Shift Review stops for a person, and the thresholds are yours to set.
Every subagent says what it will not do. 10 of them do.
What must be true in your environment
The agent can read the systems your records already live in. This one reads 14.
Somebody owns Hold Decisions and has time for it.
Somebody owns Shift Review and has time for it.

Failure Modes

What breaks this pattern:

✗ Blurred Footage Gets Judged Anyway

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.

✗ Low Confidence Graded as Fact

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.

✗ Lost Tray, Wrong Identity

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.

✗ A Score Releases the Tray

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.

Trays filmed
Counted at Line Capture
412,000
Trays judged off bad footage
Counted at Line Capture
0
Held clips handed to a person
Counted at Line Capture
41
Trays judged
Counted at Fault Finding
391,000
Faults graded by guess
Counted at Fault Finding
0
Marks held for confirmation
Counted at Fault Finding
214
Faults checked frame by frame
Counted at Tray Tracking
3,900
Single-frame marks passed on as faults
Counted at Tray Tracking
0
Line drifts raised to a person
Counted at Tray Tracking
26
Our measurement policy: We do not publish precise ROI without baseline methodology. Every figure above carries its basis.

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.

1Line Capture
subagentline-capture
writesvision/footage/<lot-id>/<tray-id>.mp4
raiseshold-bad-footage
may touchvision/footage/**, vision/held/**, the camera streams read-only
2Fault Finding
subagentfind-and-grade
writesvision/findings/<tray-id>.json
raisesbelow-confidence-floor
may touchvision/findings/**, vision/footage/** read-only
3Tray Tracking
subagenttrack-and-confirm
writesvision/tracks/<tray-id>.json
raisesfault-unconfirmed
may touchvision/tracks/**, vision/findings/** read-only
4Hold Decisions
GATE
subagentalert-and-case
writesvision/alerts/<alert-id>.json, vision/cases/<tray-id>.md
raisesawait-your-call
may touchvision/alerts/**, vision/cases/**, everything else read-only
5Shift Review
GATE
subagentclose-and-tune
writesrules/proposed/<change-id>.json, reports/line-<date>.json
raisesaccept-threshold-change
may touchrules/proposed/**, reports/**, the shift history read-only

Stack

Every system this agent reads or writes.

System
Read at
Stations
the alert rule book
Hold Decisions
2 of 5
the case file
Hold Decisions
1 of 5
the fault catalogue
Fault Finding
1 of 5
the footage store
Line Capture
2 of 5
the four line cameras
Line Capture
1 of 5
the lot record
Line Capture
2 of 5
the model's training set
Shift Review
1 of 5
the quality dashboard
Shift Review
1 of 5
the run plan
Tray Tracking
1 of 5
the severity scale the quality team wrote
Fault Finding
1 of 5
the shift record
Shift Review
1 of 5
the station lead's screen
Hold Decisions
1 of 5
the tracking record
Tray Tracking
1 of 5
thirty runs of line history
Tray Tracking
1 of 5
Next Step

Want to see if this pattern fits your computer vision?

No build commitment·Real samples, not a demo·Estimate in writing