Home Use Cases Work Insights About Contact
Back to Our Work

How a Manufacturing Automation Company Built Part Inspection That Catches the Same Defect Coming Back

Operations3
Open the walkthrough

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.

The Transformation

How a Manufacturing Automation Company Built Part Inspection That Catches the Same Defect Coming Back

Before
A person used to pull a handful of parts an hour off the belt for checking, and everything else shipped unlooked-at.
A person used to check parts at a bench, and the checking slipped as the shift wore on.
A person used to pull rejects off the belt by hand, between other jobs.
After
every part photographed, every frame clean or re-shot
every part called pass or reject, with its frames, or sent to the tray
pallet, bin or tray, each landing confirmed and counted
five parts, their cases built, waiting on your call
the running counts, the alert, and the stop that waits for you
Executive Summary

Every part that comes off the line, from the moment it passes the cameras to the moment it lands on the pallet, in the reject bin or on the hold tray. The agents photograph every part, grade every surface, place each one where its verdict says, watch the shift for a defect coming back part after part, and hand a person only the parts nobody can call, a line stopped on a defect run, and the rules for what counts as a defect. Six stations, run as a loop: what the tray corrects today is what the grader checks against tomorrow.

What Was Broken

A person used to pull a handful of parts an hour off the belt for checking, and everything else shipped unlooked-at
A person used to check parts at a bench, and the checking slipped as the shift wore on
The real cost
A person used to pull rejects off the belt by hand, between other jobs.

What We Built

Six stations and 12 subagents. Each subagent carries its own tasks and its own refusal.

A4
Camera Sweep
Photograph each part from four angles as it passes the tunnel
A3
Frame Check
Check every frame against the sharpness and lighting floor before it leaves the tunnel
A4
Part Finder
Locate each part in its four frames and separate it from the belt, the shadow and the part behind it
A3
Defect Grader
Check every surface against the defect catalogue: chips, scratches, short fills and stains
A4
Arm Guide
Compute the pick point from each part's position and turn on the belt
A3
Placement Check
Confirm each part landed where its verdict says, off the drop-zone camera
A1
Case Builder
Draft a recommendation for each tray part: keep, scrap, or send it back through the tunnel
A3
Decision Record
Record each keep-or-scrap against the frames it was decided on
A4
Pattern Watch
Count every reject by defect, spot and press across the running shift
A3
Alert Desk
Send each run to the line lead's phone with the parts and the frames behind it
A4
Shift Ledger
Reconcile every part that entered the tunnel against a landing on the pallet, in the bin or on the tray
A2
Threshold Proposal
Propose a grading change wherever the tray keeps correcting the same call

How It Runs

1

Part Capture

Nobody is asked anything here

Agents photograph every part from four angles as it passes the tunnel, match each frame to its part off the belt encoder, and re-shoot any frame too blurred or too dark to grade. Nothing waits on a person here. A part the cameras never got clean goes to the tray marked not seen, never graded off a bad picture.

2

Defect Grading

Nobody is asked anything here

Agents find the part in its frames, name which product variant it is, check every surface against the defect catalogue, and call each part pass or reject with the frames behind the call. Nothing waits on a person here. A part the grader cannot call cleanly goes to the tray with what it saw, instead of a guess going to the pallet or the bin.

3

Part Sorting

Nobody is asked anything here

Agents compute the pick point from each part's position and turn, guide the arm to place passed parts on the pallet, rejects in the bin and tray parts on the tray, and count all three against the verdicts every hour. Nothing waits on a person here. A count that disagrees with the verdicts stops the arm, and the stop reaches the line lead through the Line Watch.

4

Tray Review

A person answers here

Agents build the case for each tray part: its frames, the three most similar parts already decided, the nearest catalogue defect and a drafted recommendation. A person decides keep, scrap, or back through the tunnel. This is where a person works. Everything else exists so that five parts, not two thousand, reach this tray.

5

Line Watch

A person answers here

Agents count every reject by defect, spot and press across the running shift, send a forming run to the line lead with the parts behind it, and stop the feed when the same defect keeps coming off the same press. A stopped feed waits for a person. Agents can stop a line; only the line lead restarts one, after the press has been looked at.

6

Shift Close

A person answers here

Agents reconcile every part that entered the tunnel against the pallet, the bin or the tray, draw the shift's rejects on an outline of the part, and propose the grading changes today's tray corrections point at. A person accepts or rejects each change. What counts as a defect, and when a run stops the line, is not something an agent changes quietly.

Where a Person Decides

Step 4, Tray Review. Decide the 5 parts: keep, scrap, or send back through the tunnel. Mark any the grader read wrong; those become tomorrow's grading examples. scrap is money, a shipped bad part is more.
Step 5, Line Watch. Check press 2; the run names it, and the 7 parts sit on top of the bin. Clear the stop once the press is right; the feed will not restart itself. bad parts stop the line, a person restarts it.
Step 6, Shift Close. Accept or reject the 2 changes; nothing applies until you do. Take the shift picture to tomorrow's tool check; press 2's corner is on it. what counts as a defect is yours.

Operating Model

This changes how work flows through the team.

Role
Responsibility
Tray Review owner
Decide the 5 parts: keep, scrap, or send back through the tunnel. Mark any the grader read wrong; those become tomorrow's grading examples. scrap is money, a shipped bad part is more.
Line Watch owner
Check press 2; the run names it, and the 7 parts sit on top of the bin. Clear the stop once the press is right; the feed will not restart itself. bad parts stop the line, a person restarts it.
Shift Close owner
Accept or reject the 2 changes; nothing applies until you do. Take the shift picture to tomorrow's tool check; press 2's corner is on it. what counts as a defect is yours.
Part Capture, when it goes wrong
Look here only if the tray keeps filling with parts marked not seen.
Defect Grading, when it goes wrong
Look here only if a reject looks fine when it comes out of the bin.
Part Sorting, when it goes wrong
Look here only if the pallet count and the shipping count disagree.

What Transfers, What Must Be True

What transfers
A person is in the loop wherever a part nobody could call gets decided, a stopped line gets restarted, or the rules the agents grade by change. Three stations refuse to proceed on their own, Tray Review, Line Watch and Shift Close, and those are the three a person carries. Everywhere else the agents photograph, grade, place and count, and reach you only when the call is yours to make.
Tray Review stops for a person, and scrap is money, a shipped bad part is more.
Line Watch stops for a person, and bad parts stop the line, a person restarts it.
Shift Close stops for a person, and what counts as a defect is yours.
Every subagent says what it will not do. 12 of them do.
What must be true in your environment
The agent can read the systems your records already live in. This one reads 21.
Somebody owns Tray Review and has time for it.
Somebody owns Line Watch and has time for it.
Somebody owns Shift Close and has time for it.

Failure Modes

What breaks this pattern:

✗ Graded off a bad frame

The grader scores whatever frame it gets, sharp or not. A defect hidden by blur ships as a good part.

✗ Forced calls on unreadable parts

The grader picks a side on every part, even the ones it cannot read. A wrong guess sends a bad part to the pallet or a good part to the bin.

✗ The arm buries the miscount

The count says something is off and the arm keeps placing. Every part placed after that mixes graded parts into wrong bins, and the sort has to be redone by hand.

✗ The line restarts itself

The feed stops on a defect run and the agent restarts it before anyone checks the press. The same defect runs again at full speed into the same bins.

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.

Parts photographed
Counted at Part Capture
480,000
Parts graded off a blurred frame
Counted at Part Capture
0
Parts sent to the tray unseen
Counted at Part Capture
74
Parts graded
Counted at Defect Grading
480,000
Verdicts given without frames behind them
Counted at Defect Grading
0
Parts it would not call
Counted at Defect Grading
1,310
Parts placed
Counted at Part Sorting
480,000
Passed parts found in the reject bin
Counted at Part Sorting
0
Arm stops handed to the line lead
Counted at Part Sorting
3
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.

1Part Capture
subagentshoot-and-check
writesline/frames/<part-id>/
raisespart-not-seen
may touchline/frames/**, the camera feed read-only
2Defect Grading
subagentfind-and-grade
writesline/verdicts/<part-id>.json
raisesno-clean-call
may touchline/verdicts/**, line/frames/** read-only
3Part Sorting
subagentpick-and-confirm
writesline/placed/<part-id>.json
raisescount-mismatch
may touchline/placed/**, the arm controller, line/verdicts/** read-only
4Tray Review
GATE
subagenttray-case
writestray/decisions/<part-id>.json
raisesawait-tray-decision
may touchtray/**, line/** read-only
5Line Watch
GATE
subagentwatch-the-line
writesalerts/<alert-id>.json
raisesfeed-stop
may touchalerts/**, the feed's stop line, line/verdicts/** read-only
6Shift Close
GATE
subagentclose-the-shift
writesrules/proposed/<change-id>.json, reports/shift-<date>.json
raisesaccept-grading-change
may touchrules/proposed/**, reports/**, the shift history read-only

Stack

Every system this agent reads or writes.

System
Read at
Stations
the alert board over the line
Line Watch
1 of 6
the belt encoder
Part Capture
1 of 6
the defect catalogue
Defect Grading
2 of 6
the drawings for the nine product variants
Defect Grading
1 of 6
the drop-zone camera
Part Sorting
1 of 6
the frame store
Part Capture
3 of 6
the grading thresholds
Shift Close
1 of 6
the hold tray
Part Sorting
2 of 6
the light tunnel
Part Capture
1 of 6
the line cameras
Part Capture
1 of 6
the line lead's phone
Line Watch
1 of 6
the pallet
Part Sorting
1 of 6
the past tray decisions
Tray Review
1 of 6
the pick-and-place arm
Part Sorting
1 of 6
the presses
Line Watch
1 of 6
the reject bin
Part Sorting
1 of 6
the shift dashboard
Shift Close
1 of 6
the shift's part records
Shift Close
1 of 6
the stop rules
Shift Close
1 of 6
the verdict record
Defect Grading
1 of 6
the verdict stream
Line Watch
1 of 6
Next Step

Want to see if this pattern fits your computer vision?

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