Use casesManufacturing

Visual inspection

Inspectors check every part under a lamp; night shift misses more, and each nonconformance report takes half a day.

Kinds of AI
ModelsSLMsAgentic AI
Who decides
Quality engineer
FIG. 1 · Inspection record and draft NCRIllustrative
Part · camera 2Housing H-220

Model · anomaly gate

Rejectscore above the reject threshold

VLM · what the image and label say

  • Defect"Porosity cluster, 3–4 small pores, flange zone B, near bolt hole 2"class POR-02
  • OCRHeat no. read with high confidence
  • OCRCertificate type matches the lot record

Agent · draft NCR, not issued

  • ProcessDegassing time below spec: the likely linksame shift, die 4
  • D2POR-02 on several parts, die 4, night shift
  • ContainmentProposal: hold the lot pending a full re-check
  • Model
  • SLM
  • Agent

DecidesQuality engineer

1. The desk

Today

  1. InspectTwo inspectors check every casting by eye; consistency drops late in the shift.
  2. ReadHeat numbers, labels, and certificate fields are read and typed by hand.
  3. LinkWhen a defect is found, an engineer links it to the batch's process data by hand.
  4. WriteTogether with writing the nonconformance report (NCR), that takes the engineer half a day.
  5. DecideThe quality engineer decides whether to hold the lot only once the report is done.

With the system

  1. InspectThe anomaly model, trained on good parts only, scores every part and makes the gate call inside the quality manager's thresholds; parts in the review band go to an inspector.Model
  2. ReadThe vision-language model names the defect's class, location, and size in words, and reads the heat or lot number, markings, and certificate fields.SLM
  3. LinkThe SLM links the defect to the batch's process parameters and certificate records, and drafts the NCR.SLM
  4. WriteThe agent opens the draft NCR in the quality system, attaches the images and batch data, and proposes holding the lot.Agentic AI
  5. DecideThe quality engineer receives the draft NCR with its evidence, decides hold or release, and issues the report, as before.Person decides

What lands on the desk

The quality engineer gets the rejected part with its picture, the defect in words, the label read back, and the likely process cause, in a report that is drafted but not issued. Holding the lot is a proposal until the engineer decides.

FIG. 2 · Inspection record and draft NCRIllustrative
Part · camera 2Housing H-220

Model · anomaly gate

Rejectscore above the reject threshold

VLM · what the image and label say

  • Defect"Porosity cluster, 3–4 small pores, flange zone B, near bolt hole 2"class POR-02
  • OCRHeat no. read with high confidence
  • OCRCertificate type matches the lot record

Agent · draft NCR, not issued

  • ProcessDegassing time below spec: the likely linksame shift, die 4
  • D2POR-02 on several parts, die 4, night shift
  • ContainmentProposal: hold the lot pending a full re-check
  • flaggedQuality engineer approvalneeded
  • Model
  • SLM
  • Agent

DecidesQuality engineer

2. The system

For your technical team — skip it if you like.

How it works

In

  • Line camera images
  • Fixed lighting and fixture

  1. Model · estimates

    Image anomaly model

    Good parts only: PatchCore, PaDiM, or EfficientAD class

    Reads

    • Everything under In

    Out

    • An anomaly score per part
    • Pass, review band, or reject
    • This model makes the gate call

  2. VLM · reads images

    Defect and label reader

    Small VLM fine-tuned on the plant's defects, with OCR

    Reads

    • Flagged images
    • Every label image

    Out

    • Defect class, location, and size in words
    • Heat or lot number and markings
    • Certificate fields

  3. SLM · understands

    Report drafter

    Trained on the plant's nonconformance reports

    Reads

    • The VLM's output
    • The batch's process parameters
    • Certificate records

    Out

    • The likely link to process data
    • A draft nonconformance report (NCR)

  4. Agent · prepares

    Report-filing agent

    Fixed workflow

    Reads

    • The draft report

    Out

    • Draft NCR opened in the quality system
    • Images and batch data attached
    • Containment proposed: hold the lot

  5. Decision

    Quality engineer and inspector

    Reads

    • The draft NCR with its evidence

    Out

    • The inspector rules on the review band
    • The quality engineer decides hold or release and issues the NCR
How the agent's work is arrangedFixed workflow

Fixed workflow

The steps after a reject never change (open the report, attach the evidence, propose a hold), so a fixed sequence is enough and easier to audit.

How the models are trained and tested

Your expert writes the test first. A step the test does not need is skipped, and we say why.

Show the training steps

Anomaly model

  1. Data

    A few hundred to a few thousand images of good parts per camera view.

  2. Thresholds

    Set by weighing the cost of an escape against a false reject, with a review band for people.

  3. Validation evidence

    Escape and false-reject rates on held-out production images with labeled defects.

  4. Monitoring

    Lighting and fixture drift, after go-live.

Vision-language model (VLM)

  1. Test set

    Labeled defect images and label photos graded by your inspectors: class, location, and text-field accuracy.

  2. Dataset

    10–50 examples per defect class, and label images.

  3. Continued pre-training (CPT)Skipped

    Skipped: defect names and label fields are taught by fine-tuning; no new field language is missing.

  4. Supervised fine-tuning (SFT)

    LoRA fine-tuning (a light method that trains a small set of added weights) on (image, class, description) pairs.

  5. Reinforcement learning (RL)Skipped

    Skipped: fine-tuning on graded examples covers the task.

Small language model (SLM)

  1. Test set

    Past nonconformance reports.

  2. Supervised fine-tuning (SFT)

    On the plant's NCR format.

  3. Reinforcement learning (RL)Skipped

    Skipped: fine-tuning on the plant's reports covers the task.

Where the agentic system stops

The anomaly model makes the gate call inside thresholds the quality manager sets; the inspector rules on the review band; the quality engineer decides hold or release and issues the report.

Show its tools, limits, and record
Tools and access
Read the manufacturing execution system and certificate records; create a draft NCR; propose a hold.
May not
Use the VLM as pass or fail; scrap, release, or ship; or change the reject threshold.
Who decides and the written limit
Quality engineerThe anomaly model makes the gate call inside thresholds the quality manager sets; the inspector rules on the review band; the quality engineer decides hold or release and issues the report. The vision-language model explains and classifies; it never decides pass or fail.
Stops when
Text-reading confidence is low or the lot record is not found: the part goes to an inspector. Camera or lighting drift is detected: the gate falls back to manual inspection.
Record
The image, score, threshold version, VLM output, and the engineer's decision.

How agentic systems stay inside limits

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. Fewer defective parts escaping to the customer.
  2. Inspector time moved to the review band.
  3. Faster nonconformance reports, and labels traced on every part.

What your institution measures

  • Escape rate at the customer (parts per million)
  • False-reject rate
  • Scrap rate
  • First-pass yield
  • NCR cycle time
  • Text-field error rate on labels
Published elsewhere
On the MVTec AD benchmark (clean benchmark images, not a production line), the PatchCore method reached an image-level AUROC (area under the ROC curve, a measure of how well defective images are told apart from good ones) of up to 99.6%.

Roth et al., "Towards Total Recall in Industrial Anomaly Detection", CVPR 2022

Measured in the Decision Blueprint

Baseline escape and false-reject rates come from your inspection records; the plant sets the target.

Conditions

The rules, the data, and the known error, with its guard.

  • Lighting or fixture changes break models: drift checks.
  • New defect types are why the gate is unsupervised: it learns from good parts only.
  • Safety-critical parts need your own validation. Operators' hands and faces in images are personal data: crop or mask.
  • Model licenses (some Qwen and Gemma variants) are checked before use.

Start here

We look at your cameras, lighting, and inspection records on one line, collect good-part images, and measure today's escape and false-reject rates before a model is trained.

Request a Decision BlueprintWhat the Decision Blueprint covers
  • ModelsSLMsAgentic AI

    Maintenance on the bottleneck line

    TodayThe press fails on the night shift; its vibration had changed weeks before.

    With the systemRising stop risk flagged, with a draft work order and a proposed slot.

    Maintenance planner decides

  • SLMsAgentic AI

    Customer complaint 8D

    TodayContainment is due tomorrow, and the evidence is in five systems.

    With the systemEvidence from five systems in one 8D draft.

    Quality engineer decides

Equalize the gap.

Bring one decision. Leave knowing the first step.

A short description of the decision is enough; no confidential records.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/use-cases/visual-inspection · Printed copy; the website is the current version.