Concept buildA worked example of how we would build this for a typical business. Not client work: no client, no client data, no achieved results.
Custom AI
Concept buildA concept build for a camera inspection step that sorts machined parts into pass, flagged or unsure with a detector trained on the plant’s own images. Unsure parts always go to a person, and the quality team keeps the final say on scrap and rework.
The brief
A typical small manufacturer in India inspects parts by eye at the end of the line. Inspectors are skilled, but attention fades over a long shift, defect calls vary from person to person, and the evidence for a rejected batch is often a tally sheet rather than an image. When a customer returns a lot, the plant struggles to show what was checked and when. Off-the-shelf vision systems are usually priced and specified for high-volume lines, and general-purpose AI models know nothing about this plant’s parts or what its customers count as a defect.
Explainer
Explainer · 7 slides · about 44 s
Concept buildConcept build: Visual Defect Triage for a Small Manufacturer
Concept build, explained in 7 slides: a concept build for a camera inspection step that sorts machined parts into pass, flagged or unsure with a detector trained on the plant’s own images.
Slide 1 of 7
01 · Custom AI
Visual Defect Triage for a Small Manufacturer
Concept build: Visual Defect Triage for a Small Manufacturer.
02 · The problem
What a typical small manufacturer is up against
A typical small manufacturer in India inspects parts by eye at the end of the line.
03 · The approach
How we would build it, in 5 steps
We would build it in 5 steps, from defect catalogue and image collection to plant deployment and drift checks.
04 · Architecture
How the pieces connect
How the parts connect, from the first input to the person who reviews the result.
05 · Demo · sample run
Three parts at one inspection station
A sample run, step by step, including where a person steps in.
06 · Typical targets
What we would measure it against
Typical targets, each with the basis it would be measured against.
07 · Next step
Scope a build like this for your business
Book a strategy call to scope a build like this.
Architecture
Architecture of the Visual Defect Triage for a Small Manufacturer concept: Station camera → Image prep (photo); Labelled images → Training job (labelled set); Image prep → Defect detector (frame); Training job → Defect detector (weights); Defect detector → Threshold rules (class + score); Threshold rules → QC inspector (unsure); Threshold rules → Inspection log (pass or flag); QC inspector → Inspection log (decision); Inspection log → Flagged items view (by shift).
Diagram with 9 parts: Station camera (source), Labelled images (source), Image prep (processing), Training job (processing), Threshold rules (processing), Defect detector (model), QC inspector (human), Inspection log (system of record), Flagged items view (system of record).
Connections:
Sample run
Part photographedRoutine
Sample part: a machined bracket from batch B-12, under fixed station lighting.
Clear passRoutine
No defect above the threshold; the image and result are logged.
Defect flaggedStopped
Second part: burr on the mounting edge, high confidence, held for a rework decision.
Unsure resultNeeds care
Third part: a possible scratch, with confidence between the two thresholds.
Inspector reviewsA person takes over
The image is shown with the area marked; the inspector confirms a scratch.
Logged for retrainingRoutine
The decision is saved and the image joins the next labelling round.
Build plan
We start from the quality team’s defect catalogue: each class, an example and what the customer accepts. A fixed camera and controlled lighting go on one inspection station, specified with the plant’s maintenance team, and we collect images of good and defective parts across shifts and batches. Rare defects get a longer collection window, not synthetic images.
Quality staff label images in an open-source labelling tool, and a second inspector re-labels a sample so we can see how often people agree. Before any training, a hold-out set of the plant’s images is locked away and the recall target for each defect class is agreed against it, so the test cannot be tuned to flatter the model.
We fine-tune a permissively licensed open-weight detector, such as YOLOX under Apache-2.0, because some popular detector releases use the AGPL, which sits badly with closed plant software. Training reports recall and false rejects per class, and every model version is stored with the exact data it was trained on.
Each result carries a class and a confidence score. A confident result passes or flags the part automatically; anything between the agreed thresholds goes to a review queue, where an inspector sees the image with the suspect area marked. The inspector’s decision is logged, and the image joins the next labelling round.
Inference runs on a GPU workstation at the station or in the plant’s private cloud, so images stay under the plant’s control. A flagged-items view shows defects by shift, batch and class, and a weekly drift check compares recent results with the hold-out baseline, prompting re-labelling when lighting, tooling or material changes.
Typical targets
Typical target
80–90% recall on the plant’s known defect classes
Basis: Hold-out evaluation on the plant’s own labelled images, agreed before launch and reported per class
Typical target
Unsure parts always reviewed by a person
Basis: Design property: results between the agreed thresholds go to the review queue, never to pass or scrap
Typical target
False rejects held under a ceiling the quality team sets
Basis: Thresholds are tuned on the hold-out set; flagged parts are reviewed, not scrapped automatically
Typical target
An image and a decision on record for every inspected part
Basis: Design property: each inspection logs the image, result, confidence and any reviewer decision
Your project
Where each piece of data goes and which provider processes it; residency is checked per project, never assumed.
Candidate models compared on a held-out sample of your own documents or images before one is chosen.
On accounts in your name, self-hosted open-weight models, or on-premise as an Enterprise scope.
Drift checks, a review cadence and the steps to update the model or the index when your data changes.
Where this fits
FAQ
No plant is. This concept build describes how we would set up inspection for a typical small manufacturer in India, and the bracket, batch and defects in the demo are sample data. The recall range can only be confirmed on a real plant’s own hold-out images, which is why that test set is agreed before any training starts.
It depends on the part and the defect. A few hundred labelled examples per defect class is a common starting point; clear, frequent defects need fewer, while subtle or rare ones need more. Image collection usually sets the timeline, which is why the plan starts collecting on the first day.
No. It sorts parts into pass, flagged and unsure, and people decide what happens to anything that is not a clear pass. Scrap and rework stay quality-team decisions, recorded with the image the system flagged, which also gives the plant evidence to show a customer when a lot is questioned.
Because it decides what the plant may do with its own software. Some widely used detector releases are under the AGPL, which can bring source-sharing obligations or the cost of a commercial licence. We default to permissively licensed detectors such as YOLOX under Apache-2.0 and record the licence of every model and library in the handover.
No. Training can run on a workstation on site or in a private cloud account the plant controls, and inference runs at the station. Images stay in the plant’s own storage; the only time they move is if the plant chooses to rent cloud compute for a training run.
Worked examples
Concept builds are worked examples, not client work.
Concept buildCustom AI
A concept build for an internal assistant that answers staff questions from a firm’s own documents, inside the firm’s cloud account. Every answer cites its sources, and retrieval is filtered by each person’s existing document permissions, so nobody gets an answer built from a file they could not open.
Next step
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment
