Home Use Cases Work Insights About Contact
Back to Our Work

How a Manufacturer Built AI Vision Inspection with the Inspector Still Making the Call

Production3
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 3, 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 Manufacturer Built AI Vision Inspection with the Inspector Still Making the Call

Before
An inspector used to look at every part on the line.
An inspector used to call every defect by eye.
After
every part photographed the same way, or the station stops
every part called, with the drawing tolerance attached to the call
201 flags turned into 201 one-look decisions
the short queue of lots only a person can release
200 parts out of a 4,200 lot, checked against what the model claimed
Executive Summary

Photographs every part, calls the ones that look wrong, and hands an inspector a decision instead of a shift of scanning. Written for a precision manufacturer making machined components for aerospace and medical device customers, where the constraint was never how hard anyone was trying. A Sandia study of 82 inspectors on precision parts measured an 85% hit rate on defective items and a 35% false-alarm rate on good ones, and detection starts falling inside the first 30 minutes on task. Six stations, run as a loop: what station 6 learns from the overrides and the escapes is what station 2 calls on.

What Was Broken

An inspector used to look at every part on the line
The real cost
An inspector used to call every defect by eye.

What We Built

Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.

A4
Frame Grabbing
Capture every part from each fixed angle as it reaches the station
A3
Rig Watching
Check each fixture holds the part in the position the model was trained on
A3
Defect Calling
Score every part against the defect classes this family has shown before
A4
Threshold Holding
Track the flag rate against what this line ran the four weeks before
A2
Case Presenting
Summarize each flagged part into the one call an inspector has to make
A3
Escape Watching
Check every inspector override back against what the model actually saw
A4
Lot Grouping
Group every called part by its lot, its shift and the tool that made it
A2
Impact Drafting
Summarize each held lot into the decision a person has to make
A4
Sample Drawing
Draw the audit sample from what the model passed, at the level the plan sets
A3
Escape Counting
Check every audited part against the defect classes the model calls
A4
Override Reading
Read every override and every escape back against the call that produced it
A3
Class Watching
Search the escapes for a defect class the model has never been trained on
A1
Change Proposal
Draft the retraining this quarter's overrides and escapes support

How It Runs

1

Image Capturing

Nobody is asked anything here

Subagents photograph each part from every fixed angle, check the exposure against the enclosure's own reference, and stop a station whose lighting or fixture has drifted. Nothing waits on a person here. A part reaches an inspector only when station 3 asks them to look.

2

Defect Scoring

Nobody is asked anything here

Subagents score every part against the defect classes this family has actually shown, and lean toward flagging, because a missed defect costs more than a second look. A person sees a flag with the drawing beside it, not a verdict.

3

Flag Reviewing

A person answers here

Subagents turn each flag into a decision an inspector can make in one look, with the frame, the defect class and the drawing tolerance side by side. An inspector decides. This is the station the whole build exists to make small enough to do well.

4

Lot Deciding

A person answers here

Subagents group every called part by lot, shift and tool, count what it is worth, and draft the disposition the evidence supports. A person decides scrap, rework or use-as-is. Nothing ships or scraps on an arithmetic.

5

Sample Checking

A person answers here

Subagents draw the audit sample from the parts the model passed, check each one against the classes the model calls, and count what got past. An inspector signs the audit. The point of this station is to find out where the model is wrong, not to confirm it is right.

6

Model Retraining

A person answers here

Subagents read every override and every escape back against the call that produced it, find the defect class the model has never seen, and propose the retraining. A person accepts or rejects each proposed change. Nothing about what the model calls changes on its own.

Where a Person Decides

Step 3, Flag Reviewing. Work 201 flags, not 4,940 parts. Settle the 6 marks the drawing does not cover, with the customer if you have to. a person calls the part.
Step 4, Lot Deciding. Decide three lots, starting with the one going to the customer who publishes zero defects. Take the tool out of service, or accept the next lot from it. a person releases the lot.
Step 5, Sample Checking. Sign the audit and fail the lot, because the plan is zero acceptance. Send the escaped part to station 6 as training data. an escape fails the lot.
Step 6, Model Retraining. Accept or reject the retraining on tool marks. Take subsurface porosity to the tool owners before it becomes a defect class. nobody retrains but you.

Operating Model

This changes how work flows through the team.

Role
Responsibility
Flag Reviewing owner
Work 201 flags, not 4,940 parts. Settle the 6 marks the drawing does not cover, with the customer if you have to. a person calls the part.
Lot Deciding owner
Decide three lots, starting with the one going to the customer who publishes zero defects. Take the tool out of service, or accept the next lot from it. a person releases the lot.
Sample Checking owner
Sign the audit and fail the lot, because the plan is zero acceptance. Send the escaped part to station 6 as training data. an escape fails the lot.
Model Retraining owner
Accept or reject the retraining on tool marks. Take subsurface porosity to the tool owners before it becomes a defect class. nobody retrains but you.

What Transfers, What Must Be True

What transfers
A person is in the loop wherever a part's fate or the standard itself is decided. Four of the six stations refuse to proceed on their own, Flag Reviewing, Lot Deciding, Sample Checking and Model Retraining, and those are the four a person carries. Everywhere else the subagents run at volume and reach you only when they cannot settle something.
Flag Reviewing stops for a person, and a person calls the part.
Lot Deciding stops for a person, and a person releases the lot.
Sample Checking stops for a person, and an escape fails the lot.
Model Retraining stops for a person, and nobody retrains but you.
Every subagent says what it will not do. 13 of them do.
What must be true in your environment
The agent can read the systems your records already live in. This one reads 25.
Somebody owns Flag Reviewing and has time for it.
Somebody owns Lot Deciding and has time for it.
Somebody owns Sample Checking and has time for it.
Somebody owns Model Retraining and has time for it.

Failure Modes

What breaks this pattern:

✗ Drift becomes the new normal

When the station keeps scoring while its lighting reference drifts, every score after the drift is a score of the light, not the part. The numbers stay green while the calls go quietly wrong.

✗ The threshold walks on its own

A model allowed to move its own threshold moves it toward whatever makes today's queue quiet. Your acceptance standard changes and nobody signed the change.

✗ Lots ship past an open question

If the system releases a lot while a decision on it is still open, the parts leave the building before the answer arrives. A held pallet is cheap. A recall is not.

✗ The model learns from silence

An override with no reason written on it is a data point with no meaning. A model that trains on it learns the inspector's habit instead of the standard, and the habit spreads to every future call.

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 imaged
Counted at Image Capturing
4,940
Scored on a bad frame
Counted at Image Capturing
0
Frames refused
Counted at Image Capturing
14
Parts scored
Counted at Defect Scoring
4,940
Thresholds moved without a person
Counted at Defect Scoring
0
Declined as an untrained family
Counted at Defect Scoring
11
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.

1Image Capturing
subagentframe-grab
writeslines/<line-id>/frames/<part-id>.json
raiseslighting-out-of-reference
may touchlines/**, drawings/** read-only
2Defect Scoring
subagentdefect-score
writeslines/<line-id>/scored/<part-id>.json
raisesfamily-not-trained
may touchlines/**, drawings/** read-only
3Flag Reviewing
GATE
subagentflag-present
writeslines/<line-id>/reviewed/<part-id>.json
raisesawait-inspector-call
may touchlines/<line-id>/reviewed/**, everything else read-only
4Lot Deciding
GATE
subagentlot-brief
writeslots/<lot-id>/disposition.json
raisesawait-human-disposition
may touchlots/**, everything else read-only
5Sample Checking
GATE
subagentaudit-draw
writeslots/<lot-id>/audit.json
raisesescape-found
may touchlots/**, lines/** read-only
6Model Retraining
GATE
subagentmodel-read
writesanalytics/model/<period>.json
raisespropose-retraining
may touchanalytics/**, lines/** read-only

Stack

Every system this agent reads or writes.

System
Read at
Stations
the MES
Defect Scoring
1 of 6
the MES and ERP
Lot Deciding
1 of 6
the audit bench
Sample Checking
1 of 6
the cameras at each station
Image Capturing
1 of 6
the customer commitments
Lot Deciding
1 of 6
the customer's own acceptance requirement
Sample Checking
1 of 6
the defect class library
Defect Scoring
4 of 6
the deviation records
Lot Deciding
1 of 6
the drawing
Flag Reviewing
1 of 6
the drawing library
Image Capturing
1 of 6
the drawing tolerances
Defect Scoring
1 of 6
the escape log
Sample Checking
2 of 6
the frames
Flag Reviewing
1 of 6
the historical defect images
Defect Scoring
1 of 6
the inspector queue
Flag Reviewing
1 of 6
the lighting enclosures
Image Capturing
1 of 6
the line controller
Image Capturing
1 of 6
the line's own flag-rate history
Defect Scoring
1 of 6
the lot records
Lot Deciding
1 of 6
the override log
Flag Reviewing
2 of 6
the part fixtures
Image Capturing
1 of 6
the sampling plan
Sample Checking
1 of 6
the scrap and rework ledger
Model Retraining
1 of 6
the tool history
Lot Deciding
1 of 6
the training set
Model Retraining
1 of 6
Next Step

Want to see if this pattern fits your manufacturing quality?

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