Forward-Deployed Engineering · Automotive components · Quality inspection
Case study: visual inspection of machined castings, from lab bench to production line
Illustrative engagement — not a client record. The company, people, volumes and results below are a representative composite, written to show how a forward-deployed engagement runs end to end. The workflow, the architecture, the controls and the method are real and technically valid.
Offering usedPilot-to-ProductionA stalled pilot hardened, integrated and put in front of real users.
Illustrative results, first 90 days after go-live, in-scope part families only:
100%
Parts with an image record linked to their serial code
Before 0%
11 ppm
Customer-reported defects on in-scope features
Before 41 ppm
1.3%
Good parts held for re-inspection
Before 3.6%
under one hour
Time from a defect spike to a process engineer seeing it
Before next morning's report
Explainer09 sheets
The whole story in about a minute
Nine short scenes, from the inspection table to the results. Press play, or pick a scene.
View ABy eye
Parts at the table
About 7,000 housings a day
- Escaped
Sheet01 / 09
ClockBefore
Every part was checked by eye, under a lamp.
About 7,000 housings a day, two inspectors per shift. Defects were rare, so they were hard to see, and some still reached the customer.
- 1.About 7,000 housings a day
- 2.Two inspectors per shift
- 3.Under 2 in 100 parts had a defect
At a glance
- Client
- A tier-one supplier of aluminium die-cast and machined housings for steering and braking systems, one plant, two machining lines, IATF 16949 certified
- Workflow
- End-of-line visual inspection of machined housings: image each part, find surface defects and missing features, send it to pass, reject lane or a human inspector, and record every decision against the part's serial code
- Engagement
- Pilot-to-Production (ten weeks, plus six working days of paused clock)
- Team
- One named senior forward-deployed engineer, full time, with a second engineer for the three integration weeks. Client side: the Plant Quality Head (owner), a controls engineer, the MES analyst, two senior inspectors, an OT security reviewer
- Where it runs
- On an inspection computer at the station, inside the plant network. Images and records stay on the plant's own servers
- Handover
- Runbook executed by the client's controls engineer in week ten; our access revoked the same day
| Measure | Before | After |
|---|---|---|
| Parts with an image record linked to their serial code | 0% | 100% |
| Customer-reported defects on in-scope features | 41 ppm | 11 ppm |
| Good parts held for re-inspection | 3.6% | 1.3% |
| Time from a defect spike to a process engineer seeing it | next morning's report | under one hour |
| Parts released by the model that a rule would have held | not applicable | 0, by design |
FDE / 01Case study
01 / 12
In shortPeople checked every part by eye, and rare defects still reached the customer.
The plant machined around 7,000 aluminium housings a day across two lines and three shifts. After machining and washing, every part passed an inspection table. Two inspectors per shift turned each housing under a lamp and looked at the sealing faces, the bores, the threads and the cross-drilled holes.
They were looking for four things: porosity opened up by machining on a sealing face, cracks near the mounting bosses, burrs left inside cross-drilled holes, and missing features such as an untapped thread or a skipped hole. Most days, fewer than two parts in a hundred had any of them.
The cost showed up in three places:
Escapes. The customer returned parts with porosity on a sealing face or a burr in an oil passage several times a quarter. Each complaint meant a containment, a sort at the customer's plant and a formal corrective action report.
Traceability. When a complaint arrived, the plant could say which shift inspected the batch. It could not show what that particular part had looked like when it left.
People. Inspection is tiring and rare defects are the hardest to see. Inspectors held good parts when in doubt, and a second person re-checked them at the end of the shift.
FDE / 02Case study
02 / 12
In shortA vision pilot did well on a lab bench, but failed on the real line within two shifts.
Eighteen months earlier the plant's innovation cell had run a vision pilot. On a lab bench, with a ring light and 3,000 hand-picked images, it reported 97% accuracy. It was installed on line one for a week and switched off after two shifts. The reasons are common rather than unusual:
The bench was not the line. On the line, wet parts came straight from the washer, a roll-up door and roof lights changed the ambient light through the day, and the fixture let each part sit a little differently. In its first shift the pilot sent four in ten good parts to the reject bin.
Accuracy was the wrong number. With defects under 2% of parts, a model that passes everything is 98% accurate. Nobody had measured how many real defects it missed, by defect type.
It was not connected to anything. Decisions appeared on a screen. The reject gate, the MES and the part's serial code were all "phase two".
It had no owner. The pilot belonged to the innovation cell, not to the Quality Head. When it was switched off, nobody was accountable for switching it back on.
The model architecture was reasonable. Lighting, integration, evaluation and ownership were not.
FDE / 03Case study
03 / 12
In shortTwo weeks on the shop floor, ending in signed, measurable rules for going live.
The engineer spent the first two weeks on the shop floor: three shifts beside the inspectors, two days with the controls engineer at the line PLC, and time with the MES analyst and the quality engineers who handle complaints.
What was found:
FIG. 3.1- Lighting was the real work. A test with a dark cloth over the station showed that most of the pilot's false rejects came from reflections and water spots, not from the model. The station needed an enclosure, a dome light for the faces, low-angle light for burrs and cracks, and an air knife to dry the part.
- A false assumption. Everyone believed the part's laser-marked serial code was read at inspection. It was read only at packing. Without a code reader at the station, no image could be tied to a part.
- Two foundries. Castings came from the plant's own foundry and from an outside supplier. The supplier's parts were shot-blasted and looked different on unmachined surfaces.
- The defect catalogue existed. The customer had approved limit samples for each defect type: the largest pore, the smallest burr, that could be accepted. That catalogue became the backbone of the evaluation.
The gap report and go-live criteria, signed by the Plant Quality Head:
- WorkflowEnd-of-line inspection of the two highest-volume housing families on both lines: machined faces, bores, threads and cross-drilled holes
- Out of scopeUnmachined as-cast cosmetic surfaces, internal porosity that does not reach a surface, dimensional measurement
- MetricZero misses on the seeded critical defects in the golden set, and a false-hold rate below 2% of good parts
- OwnerPlant Quality Head
- GuardrailsThe model never releases a part a deterministic rule would hold. Stopping the line stays with the shift supervisor. Inspectors audit a sample of passed parts every shift
- Go-live criteriaConnected to the PLC and the MES; scored on the golden set; OT security review signed; monitoring live with a named person paged; rollback rehearsed; controls engineer trained
Timeline09 stations
The engagement, stage by stage
Nine stages on one clock, from the bench pilot to ninety days after go-live. Press play, or pick a stage.
LineStopped
StationOP 10 By eye
ClockBefore
Work instructionOP 10 · 01 / 09
Every part was checked by eye, under a lamp.
About 7,000 housings a day, two inspectors per shift. Defects were rare, so they were hard to see, and some still reached the customer.
- About 7,000 housings a day
- Two inspectors per shift
- Under 2 in 100 parts had a defect
FDE / 04Case study
04 / 12
In shortTen weeks, plus six days paused for the enclosure: light it, connect it, test it, switch it on, hand it over.
plus six working days of paused clock
W1Assess
What happened
Engineer on all three shifts; pilot code and images read; access requested for the PLC, MES and plant servers. First pull request merged on day seven: the pilot's code moved into the client's source control with its first tests
What existed at the end
A first gap report and a list of what to keep: the pilot's defect classes and its image labelling tool
What happened
Engineer on all three shifts; pilot code and images read; access requested for the PLC, MES and plant servers. First pull request merged on day seven: the pilot's code moved into the client's source control with its first tests
What existed at the end
A first gap report and a list of what to keep: the pilot's defect classes and its image labelling tool
What happened
Go-live criteria drafted and signed. Enclosure and lighting specified with the controls engineer
What existed at the end
Signed criteria, a lighting specification, a code reader ordered
What happened
Second engineer joined. The PLC trigger and image capture merged, writing to plant storage. The enclosure fabrication slipped; the clock paused for six working days, written down that day
What existed at the end
A capture service running on line one without lights, for plumbing only
What happened
Enclosure installed. Two weeks of real images collected on both lines. Golden set agreed with the senior inspectors
What existed at the end
A golden set of 9,400 part images with known answers
What happened
Serial code, MES record and reject-lane signal connected. OT security review of the design
What existed at the end
Every captured image tied to a serial and a record in the MES
What happened
Models trained on line images. Shadow run began: the system decided every part while inspectors worked as before. Drift monitor caught the outside foundry's new batch
What existed at the end
A system deciding in parallel with the inspectors, with dashboards
What happened
Golden set extended with the supplier's parts; thresholds set with the Quality Head; rollback rehearsed on line one
What existed at the end
A release that met every scoring criterion
What happened
Second week of shadow run; disagreements reviewed daily with inspectors
What existed at the end
A signed shadow-run comparison
What happened
Line one live, then line two four days later, each gated on the shadow comparison
What existed at the end
Real parts routed by the system, rollback one switch away
What happened
Controls engineer performed a release, a rollback and a threshold change without us at the keyboard. Criteria ticked
What existed at the end
Signed runbook, decision record, revoked access
plus six working days of paused clock
Calendar time was ten weeks and six working days. The six days were the enclosure. They were recorded when they happened, not explained at the end.
FDE / 05Case study
05 / 12
In shortA station that images every part, checks rules before models, and sends doubts to a person.
Try a part07 ops
Pick a part. Watch where the station sends it.
The same station checks every housing. What the images show decides whether it passes, goes to the reject lane or goes to a person.
Pass
Passed
The part moves on, and its images stay linked to its serial code.
Capture. When the PLC signals that a part is clamped, four cameras image it under a dome light and a low-angle ring, with the part rotated to show each face. The enclosure blocks ambient light, and an air knife dries the part first. Lighting and optics were designed with the controls engineer; the software cannot fix an image the light did not produce.
Image checks. Deterministic checks run before any model sees the image: exposure within limits, focus score above threshold, part present and seated correctly, serial code read. If any check fails, the part goes to a human inspector. A bad picture is never a pass.
Feature rules. Classical vision rules, not a learned model, confirm that each hole and thread exists where the drawing says it should. A missing feature is a hold, whatever any model says.
Defect models. Two models run side by side. A supervised detector finds the known defect types in the machined zones: porosity, cracks, burrs. An anomaly model, trained only on good parts, scores how unusual each zone looks. It exists for the defect nobody has labelled yet.
Decide. Three outcomes only:
Pass, when every image check and feature rule passes, the detector finds nothing above its threshold and the anomaly score is below its threshold.
Human inspector, with the images, the zone and the reason, when anything is uncertain: a score in the grey band, an anomaly with no known defect type, an image check failure, or a casting lot from a supplier not yet in the golden set.
Reject lane, when the detector finds a known defect in a critical zone above its reject threshold. Rejected parts are quarantined. A quality engineer decides whether each is scrap, rework or a false reject.
The order matters. Rules are evaluated first. A model score can move a part from pass towards the inspector or the reject lane. It can never move a part a rule holds towards pass.
Records. For every part the MES receives the serial code, the images, each rule result, each score, the model version and the outcome, including what an inspector later decided. Records follow the plant's existing retention policy for quality records.
Trends. Every hour, rejects and holds are grouped by zone, die, cavity and shift. When porosity on one sealing face rises for one die, the process engineer sees it that hour, not in the next morning's report.
Line stop. The system never stops the line. When three rejects of the same type arrive within twenty parts, it alerts the shift supervisor with the images. The supervisor decides.
System map20 objects
How the pieces connect
The station, the rules, the models and the people around them, lit one scene at a time. Press play, or pick an event.
Graph01 / 09
Description
Every part was checked by eye, under a lamp.
About 7,000 housings a day, two inspectors per shift. Defects were rare, so they were hard to see, and some still reached the customer.
FDE / 06Case study
06 / 12
In shortA test of 9,400 part images that every release must pass with zero missed seeded defects.
The golden set was the most important thing the engagement produced.
9,400 part images from both lines, all three shifts, both foundries and both housing families, captured under the new enclosure. Nothing from the lab bench was reused.
Answers from two senior inspectors, labelling independently, with a quality engineer settling every disagreement against the customer's limit samples.
Seeded critical defects. Forty physical parts with known porosity, cracks, burrs and missing features, each imaged many times in different positions. The criterion was zero misses. The same parts now run through the station at the start of every shift, as a daily check that the inspection still works.
Scored by defect type, not by accuracy. For each type the harness reports how many real defects were caught and how many good parts were held. With defects this rare, a single accuracy figure hides exactly the misses that matter.
An agreement study. The system's decisions were compared with the inspectors' on the same parts, using the same attribute agreement method the plant already used for its human inspectors.
The suite runs in the client's CI. A release that misses any seeded defect, or that raises false holds above the threshold, fails the build.
- IF misses any seeded defectBLOCKS
- IF raises false holds above the thresholdBLOCKS
fails the build
It earned its place in week six. During the shadow run, the drift monitor flagged that anomaly scores had risen on parts from the outside foundry. A new batch had a coarser shot-blast finish. The anomaly model was treating texture as a defect, and false holds on those parts had tripled. Nobody retrained anything quietly. The golden set was extended with 600 images from that supplier, the thresholds were re-scored, and until the new release passed, every part from that supplier's lots went to a human inspector by rule.
FDE / 07Case study
07 / 12
In shortEverything stays in the plant, and models flag while rules and people decide.
Perimeter. Inference runs on an inspection computer at the station. Images and records stay on the plant's servers. Training runs on a plant server with a GPU. Nothing leaves the site.
Network. The inspection computer sits in the plant's OT zone. It reads one PLC signal and writes one: the lane decision. It has no write access to any other controller. MES writes go through the MES's own interface as a service account.
Can
- reads one PLC signal
- writes one: the lane decision
Cannot
- write access to any other controller
Identity. Our engineers' access went through the client's identity provider and a jump host, appeared in their audit log like anyone else's, and was revoked at handover.
Autonomy is bounded. Models flag; rules and people decide. A model never releases a part a rule holds. Line stops stay with the shift supervisor. Scrap decisions stay with a quality engineer. Inspectors audit one passed part in every two hundred, and every first part after a changeover or a die change.
Review. The OT security reviewer approved the design in week five and the deployment in week eight. Findings and how each was closed are in the decision record.
FDE / 08Case study
08 / 12
In shortA two-week shadow run, then one line at a time, with one switch back to the manual table.
The shadow run was the gate. For two weeks the system decided every part while the inspectors worked as before, and each disagreement was reviewed the next morning with the senior inspectors. Disagreements fell into three piles: the system was wrong, the inspector was wrong, and the limit sample was unclear. The third pile went to the Quality Head and the customer's quality contact as questions about the catalogue.
For two weeks
the system was wrong
the inspector was wrong
the limit sample was unclear
went to the Quality Head and the customer's quality contact as questions about the catalogue
Several parts in the second pile were real burrs the inspectors had missed. Those images went into the golden set too.
Go-live went by line: line one first, line two four days later, each after a clean day of seeded-part checks and audit samples. During the first two weeks inspectors stayed at the table and re-checked every pass on a sample of shifts. Rollback was one switch that sent every part back to the manual table. It was rehearsed in week seven and never needed.
01line one first
02line two four days later
EACH STEP · a clean day of seeded-part checks and audit samples
FDE / 09Case study
09 / 12
In shortThe client's controls engineer proved they could run it before we left.
The engagement ended on the go-live criteria, not on the calendar.
In the dry run, the controls engineer added images from a new die, scored a retrained model against the golden set, released it, and rolled it back. Our engineer was in the room but not at the keyboard. The runbook was signed after that, not before.
Key numbers09 gauges
The story in nine numbers
One number for each scene, from 41 ppm of defects reaching the customer to 11. Press play, or pick a number.
ClusterStopped
ChannelOP 10 By eye
ClockBefore
DialOP 10
41
ClockBefore
Face01 / 09
Three shifts
ReadoutOP 10
Every part was checked by eye, under a lamp.
About 7,000 housings a day, two inspectors per shift. Defects were rare, so they were hard to see, and some still reached the customer.
Signals
About 7,000 housings a dayTwo inspectors per shiftUnder 2 in 100 parts had a defectFDE / 10Case study
10 / 12
In shortFewer defects reaching the customer, fewer good parts held, and a picture of every part.
Figures are illustrative, measured on the in-scope housing families over the first 90 days after go-live.
Every part has an image record tied to its serial code. A complaint now starts with the pictures of the part that left, not with a shift roster.
BEFORE0%AFTER100%Customer-reported defects on in-scope features fell from 41 ppm to 11. Most of the gain was burrs in cross-drilled holes, which people found hard to see and the low-angle light shows clearly.
Good parts held for re-inspection fell from 3.6% to 1.3%. Holds now arrive with the zone and the image, so re-inspection takes seconds.
BEFORE3.6%AFTER1.3%1%4%−2.3 ptsDefect spikes reach process engineers within the hour. Twice in the first quarter, rising porosity on one die was traced to a vacuum fault in the foundry the same shift.
No part released against a rule, because the design does not allow it.
No inspectors were removed. They now review held parts, run the seeded checks and audit passes. One senior inspector owns the defect catalogue.
What did not improve, and was never promised:
FIG. 10.2Porosity below the surface is still invisible to a camera. The leak test and the sampled X-ray remain exactly as they were.
As-cast cosmetic surfaces are still inspected by eye. They were out of scope, and the gap report said so.
Parts from the outside foundry still go to a human more often than the plant's own castings. The supplier's finish varies by batch, and no model change removed that.
FDE / 11Case study
11 / 12
In shortSix rules for anyone putting vision inspection on a real line.
- Fix the light before you train anything.
Most of the stalled pilot's errors were reflections. An enclosure did more than any model later would.
- Never report accuracy on rare defects.
Count misses by defect type and good parts held. Nothing else tells you whether the line is safe.
- Collect images on the real line.
A lab bench teaches the model the lab.
- Plan for a new supplier, a new die and a dirty lens.
Drift is not an if. Decide in advance who sees it and what the rule does meanwhile.
- Keep rules in front of models.
It is what made the security review and the customer's questions short.
- Pause the clock in writing.
A late enclosure is not a slip if both sides recorded it the day it happened.
FDE / 12Case study
12 / 12
In shortThe client took it in-house, extended it themselves, and came back for a leak-test station.
The client took the system in-house and did not buy managed operations. Their controls engineer extended it to a third housing family using the runbook and the harness. They returned for a Production Sprint on a leak-test station nearby, where the question is different: reading pressure-decay curves rather than images.