Physical AI
Computer vision development services, built to run where the cameras are
Our computer vision development services take a vision problem from sample images to a commissioned system: data capture, model development, optimisation for edge hardware, and integration with the line. Vision systems that make a call in milliseconds, on hardware bolted to the line. Sending video to a data centre is too slow for a reject gate, too expensive at scale, and in plenty of buildings simply not allowed.
Inspection cell 01
- p95 latency
- 13.05 ms
- Test images
- 5,000
One line, instrumented
What's already there
Cameras recording for insurance that nobody watches live. A historian holding six years of vibration readings. A quality sheet on paper that never reaches process engineering. Most of what a system needs is already on site. It just isn't connected to anything that can act.
Where it has to run
A reject gate has milliseconds. The footage isn't allowed to leave the building. The uplink will drop on the third shift. None of that is an edge case; it is the job, and it rules out anything that needs a round trip to a data centre.
What working means
Before any code, working becomes a number. Detection rate on a held-out set. Decision latency at the camera. Availability across sites. Written down, signed, and then measured in shadow mode before the system is allowed to act on anything.
What gets assembled
Five practices, one team.
And it keeps running
Products change. Lighting gets replaced. A camera moves during maintenance. Accuracy decays without announcing itself, so the loop comes back: monitored against sampled ground truth, retrained on a schedule, rolled out in waves.
What working means here
Signed first. Measured after.
The value of each one is set with you during scoping, from your data and your line. What does not change is that they are written down first and published against afterwards, pass or fail.
- 01
Detection rate
On a held-out set from your own footage, across shifts and changeovers
- 02
Decision latency
Frame in to signal out, on the device installed at the line
- 03
False-positive rate
Measured in shadow mode before the system is allowed to act
We publish the measurements behind this practice, including the runs that failed.
Vision Detection Deployment Benchmark 2026How it is assembled
Stage by stage
Each one ends in something you keep.
- Step 01
Site survey
We come to the floor. Lighting, angles, optics, mounting, network and power are measured before anything is promised.
- Step 02
Capture and annotate
Real production recorded across shifts and changeovers, labelled against the classes your quality team already uses.
- Step 03
Build and tune
Model development and optimisation until it holds the required frame rate on the hardware being installed.
- Step 04
Commission on the line
Hardware mounted, integrated with the control system, running in shadow mode alongside your existing process.
- Step 05
Shadow run and sign-off
A monitored production period reviewed against the agreed thresholds before the system acts on its own.
Scope
What the work includes
- Site survey: lighting, camera placement, optics, mounting, network and power
- Data capture and annotation, including the failure modes that matter
- Model development against your footage and your defect classes
- Optimisation to hold the frame rate your process actually needs
- Deployment to industrial edge hardware, sealed and mounted
- Integration with reject mechanisms, HMIs and existing camera systems
- Shadow-mode running before anything acts automatically
- Drift monitoring and a scheduled retraining path
Deliverables
What you keep
- 01Site survey and camera plan
- 02Annotated dataset, owned by you
- 03Validated model with a published confusion matrix
- 04Commissioned edge hardware on the line
- 05Integration with your control system
- 06Shadow-run report and sign-off
Typical stack
PyTorch · ONNX Runtime · TensorRT · OpenCV · GStreamer · NVIDIA Jetson · GigE Vision / RTSP · OPC UA
What our computer vision development services include
Each engagement covers the whole path from feasibility to a system running in production, so nothing is handed over half-finished:
- Feasibility and consulting. Sample images are tested before any build, and you get an accuracy band and hardware class in writing.
- Data capture and annotation. Footage is recorded on your floor across shifts and changeovers, and labelled against your own defect or event classes.
- Custom model development. Detection, classification, segmentation, anomaly detection and OCR models trained on that data, not a generic library.
- Optimisation for the target hardware. Models are compressed and compiled until they hold the frame rate the process needs on the edge device being installed.
- Integration. Results reach the PLC, reject gate, HMI, MES or video management system that has to act on them.
- Monitoring and retraining. Drift is tracked in production, and a retraining path is part of the handover.
Where computer vision pays back
We concentrate on the physical world, where a camera decision changes what happens next:
- Visual quality inspection of parts, surfaces and assemblies.
- Label, barcode and printed-code verification on packaging lines.
- Safety monitoring: exclusion zones, PPE and occupancy on existing cameras.
- Counting, tracking and flow measurement in production and logistics.
- Video analytics on CCTV estates that are recording but not watched.
What makes computer vision hard in production
Most computer vision projects do not fail in the model. They fail on the floor. Lighting changes between shifts, parts vary more than the sample set did, the critical defect is rare, and the decision has to arrive before the part leaves the station. A model that looked accurate on clean photographs can collapse on third-shift footage.
That is why we measure on your own footage, set acceptance thresholds before the build, and run in shadow mode before a system is allowed to act. We publish how we measure, too: in our vision deployment benchmark, a YOLO11n detector on a Cloud Run L4 GPU measured 13.05 ms at p95 across all 5,000 COCO validation images (our own measurement, in the published benchmark report). Latency of that order is why the decision can be made at the camera.
How to choose a computer vision development company
Ask these questions of any vendor, including us:
- Will you test feasibility on our images before quoting? A written accuracy band beats a slide deck.
- Will you measure on our footage, across shifts, rather than on a public dataset?
- What detection, false-positive and latency thresholds will you sign before build?
- Can it run on site without sending video out, if our compliance team requires that?
- Who owns the dataset, the trained model and the code after handover?
- How will the model be monitored and retrained when products or lighting change?
Sizing
The shapes it comes in
- Single station
- One camera, one industrial edge unit
- One inspection point, one defect family, one line
- Line
- Three to eight cameras with local aggregation
- Full line coverage with shared reporting
- Multi-site
- Managed estate across plants
- Central model management, staged rollout, per-site validation
Manufacturing · Vision
Surface defect inspection, decided at the camera
Final inspection was manual and sampled, so surface defects were being found by the customer rather than on the floor. We surveyed the cell, captured production across three shifts, and commissioned an edge vision station that inspects every part and drives the existing reject gate.
Read the engagement- On-device
- decision made at the camera, before the part leaves the cell
- Every part
- inspected, replacing sampled manual checks
- None
- imagery leaving the plant network
Start
How Edge AI & Computer Vision begins
Vision system, single line or cell: $60K–150K
Where an engagement lands inside its band is set by data readiness, integration depth, the reliability bar, evaluation burden, deployment constraints and usage economics. All six, explained.
Thirty minutes on the problem and where it happens, then a fixed-price pilot if it looks like a fit.
Arrange a site surveyFAQ
Asked before signing
Only if you want it to. The default keeps frames on the edge device, sends structured results to your systems, and works with the site's uplink disconnected.
That depends on your parts, your defects and your lighting, which is why we will not quote a number before the survey. Send sample images and the Vision Feasibility Check gives you an accuracy band in writing, free.
Yes, we can often use the cameras you already have. The survey checks whether their optics and placement can hold the accuracy you need, and says plainly when they cannot.
The model is retrained on the new data. New products and tooling changes shift what the camera sees, so engagements include a retraining path, and drift monitoring that flags a refresh before quality reports do.
A vision system for a single line or cell is indicatively $60K–150K, including hardware and commissioning, as on our pricing page. The range depends on the number of cameras, how hard the defects are to see, and how many plant systems the results feed. The feasibility check comes first, so the quote rests on your images.
A single line or cell typically runs 8–14 weeks from survey to sign-off, per our pricing page, including a shadow-mode period before the system acts. The free feasibility check before it returns a written memo in a few days.
It depends on latency, bandwidth and what the site allows to leave the building. A reject decision measured in milliseconds, or footage that cannot leave the facility, points to the edge. Batch analysis of recorded video can run in the cloud. The survey settles which, once latency needs and site rules are known.
Yes, we work with plants outside New Jersey, across the United States. We are based in New Jersey, and surveys and commissioning happen on site. To start, send sample images for a free feasibility check.
Often paired with
Where this fits with the rest
- 01Physical AI
Vision Systems Integration
Cameras, optics, lighting, enclosures, and the PLC, SCADA and MES interfaces that make results act.
- 02Physical AI
AI IoT Development
Sensor fusion, predictive maintenance and condition monitoring across a fleet you can update.
- 03Start here
Vision Feasibility Check
Send 30–50 sample images. Get a written memo: accuracy band, hardware class, risks, go or no-go.