For pixels that need to become a decision, not just a picture
Computer Vision Development
A camera captures more information than any form or ticket, yet most of it goes unused without a system built to read it. This practice turns image and video streams into structured decisions: a defect flagged, an object counted, a document captured accurately. Vision is engineered for a messy real site, not a curated dataset.
How a production vision system actually gets built
Model selection, inference deployment and the drift monitoring a live camera feed needs.
90 secSample frame · video pending01Before any model is trained
The decision a vision project cannot skip
Vision projects fail most often on data and conditions, not on model choice, so the decision belongs before training starts. Four questions settle whether a production vision system is the right constraint to solve for.
Do you have labelled examples of the target?
A model learns from examples of exactly what it should catch. Without labelled defects or objects, the first phase is collecting them, not training.
How much does the real environment vary?
Lighting, angle and motion in a live site rarely match a clean dataset. Underestimating that variance is the most common cause of a model that fails on deployment.
Does the decision need to happen in real time?
Flagging a defect as it passes a camera needs low-latency inference near the line. Reviewing footage after the fact tolerates a slower, cheaper pipeline.
What is the cost of a false accept against a false reject?
Those costs are rarely equal, and the threshold that balances them is a business decision this practice cannot make for you alone.
| Manual review | Automated vision | |
|---|---|---|
| Consistency | Varies by inspector and shift | Consistent against a set threshold |
| Throughput | Bound by headcount | Scales with camera and compute |
| Auditability | Rarely logged in detail | Every decision is inspectable |
| Setup effort | Minimal | Requires labelled training data |
| Right for | Low volume, high judgement | High volume, defined criteria |
02Agreed before any build
The ledger that keeps the scope honest
The engagement draws four lines before a model gets trained. You supply the imagery, this practice builds the pipeline, you take delivery of the system, and certain decisions never move away from you. Every scenario positions those lines a little differently, based on where your production risk actually sits.
- Your side
You bring
- Sample images of passing and failing units
- The inspection criteria your team applies today
- Camera and lighting conditions on the line itself
- Our side
We build
- A detection model trained on your own defect classes
- An inference pipeline sized to your line speed
- A confidence threshold tuned against false rejects
- Your side
You receive
- A vision system flagging defects at line speed
- A labelled dataset your team can keep extending
- Drift monitoring that catches accuracy decay early
Never leaves you
- Final sign-off on any rejected unit
- Ownership of every training image and label
- The quality policy the model is built to enforce
Where the line sitsThe model enforces a quality policy your team sets; it never gains authority to redefine what counts as a defect.
03What actually drives the number
Four drivers, not a flat rate card
Vision work is priced against the assessment, not a template, because these four factors change the effort by multiples.
Labelled data availability
Existing labelled images accelerate training significantly. Starting from raw, unlabelled footage adds a data collection and annotation phase most estimates miss.
Environmental variance on site
A controlled indoor setup with fixed lighting is simpler than an outdoor site with variable weather and motion the model must tolerate.
Real-time inference requirements
Batch processing footage overnight is far cheaper to build than a low-latency pipeline flagging defects as a line moves.
Integration into existing systems
A standalone dashboard is simple. Feeding detections into an existing quality or inventory system multiplies the integration work involved.
04Evidence and proof
Vision systems judged on catch rate, not demo footage.
Vision proof is measured accuracy on real site conditions, not a benchmark screenshot. Every record here waits for written client approval before production.
Explore case studiesA production line caught defects a shift-based inspection missed.
Nexa Foundry relied on manual spot checks that varied by shift. A trained detection model now flags defects consistently at line speed, with every decision logged.
- Defects caught before shipment
- A clear majority
- Inspection decisions logged
- Every one
Parcel damage stopped being discovered by the customer first.
Tern handled damage claims after delivery, once the customer had already complained. A vision system now flags visible damage at the hub before a parcel ships.
- Damage caught before shipment
- Substantially more
- Claims traced to hub footage
- Every flagged case
Seeing exactly which conditions confused the model early meant we fixed the camera setup instead of blaming the algorithm.
Sample · pending client approval
05Read this first
When a custom vision build is not the right fit
Selling a custom detection model to a team whose problem is really a data or process gap would cost more than it earns us in trust. The exits are marked before the pitch. When a build is justified, it runs on trained detection models, a monitored inference pipeline and the storage layer your data already lives in.
- Machine Learning Development
The real gap is a broader predictive model
If the target output is a forecast or a risk score rather than something read from an image, a different practice fits the problem better.
- AI Integration Services
The vision model needs to plug into other systems
If detections need to trigger workflows across multiple platforms, the integration layer matters as much as the model itself.
- Managed AI Services
You need the model operated, not just built
Ongoing monitoring, retraining and drift management can run as a standing service once the initial model is proven.
| Shape | Best when | You direct | We carry |
|---|---|---|---|
| Scoped build | The target classes are defined | Threshold and priorities | Model and pipeline delivery |
| Delivery pod | The scope will grow with results | Priorities each cycle | Engineering, quality, releases |
| Team extension | Your leads run delivery well | Architecture and standards | Named engineers and continuity |
06Straight answers
Asked before a vision budget commits.
01Do you run inference at the edge or in the cloud?
Whichever the latency and connectivity requirements demand. A production line needing an immediate flag usually runs inference at the edge; footage reviewed after the fact can run centrally at lower cost.
02How much labelled data do we need to start?
Enough examples of every target class to train reliably, which the assessment quantifies against your specific defect or object types rather than a generic rule of thumb.
03Does computer vision include OCR and document capture?
Yes, document and text-in-image capture sits inside this practice, engineered the same way as object detection: trained on your document formats and validated against real scan conditions.
04Who owns the training images once the model exists?
You do, along with every label and the trained model itself. This practice builds the pipeline; the data and its ownership never move away from your organisation.
05How do you handle privacy in video-based systems?
Retention, access and redaction decisions are made with your team against your own compliance duties, since privacy obligations vary by sector and by what the footage contains.
06Can an existing model be improved instead of rebuilt?
Often, yes. An assessment of the existing model, its training data and its failure cases usually finds targeted fixes before a rebuild is ever justified.
07What cameras or hardware does a vision system need?
Requirements depend on lighting, distance and required accuracy, which the assessment specifies against your site rather than a fixed hardware list sold in advance.
08How often does the model need retraining?
Whenever drift monitoring shows accuracy decaying against new conditions, which is why the operating model handed over includes that monitoring from day one.