Artificial Intelligence
The data does not exist. The imaging is wrong. The latency budget will not close. Or the system drifts and nobody is watching the right number. Our AI practice is organised around those four failures, in that order, because that is the order they actually kill programmes.
Every one of these is cheaper to discover in week two than in month seven, which is what a feasibility study is actually for.
The defect the customer cares about has occurred eleven times in two years, and nine of those were not photographed. Supervised classification is the wrong frame; anomaly detection, augmentation or better instrumentation will serve better.
How engagements are scoped → FAILURE 02The feature is not visible under the lighting that was installed, so no model can find it. Most failed inspection pilots are illumination failures wearing an algorithm costume — and the fix is a lamp, not a network.
Computer vision engineering → FAILURE 03Inference takes 80 ms and the line gives it 25. Framework operation counts correlate poorly with wall-clock latency on embedded targets, because the bottleneck is memory bandwidth — so it gets measured on the actual hardware, at the ninety-ninth percentile.
Edge AI and embedded ML → FAILURE 04A supplier changes, a lamp dims over eighteen months, a season turns. Accuracy decays across a quarter while the dashboard stays green and the model keeps returning confident answers that are no longer warranted.
MLOps for devices →Work done with your team, on your product. Feasibility first, and the feasibility study is allowed to conclude that it will not work as posed.
Feasibility, data strategy, model development, deployment, and the support that keeps accuracy from drifting away.
SERVICEQuantisation, pruning, operator fusion and accelerator mapping so a model runs in milliwatts rather than watts.
SERVICEImaging chain, illumination, detection, segmentation, OCR and OCV — engineered as a system, not a model.
SERVICECamera, LiDAR, radar and inertial aligned in time and space, then fused. Mostly a timing problem.
SERVICEAccelerator architecture, memory hierarchy, quantisation strategy, RTL and SoC integration.
SERVICEVersioning, staged over-the-air model delivery, drift detection and a retraining path you can run.
The engineering has been done before, which mostly means the expensive surprises have already been found.
Surface, dimensional, presence, print and OCV inspection with PLC integration and reject actuation.
Vibration, thermal and motor-current analytics with lead time you can schedule around.
Infrared in-cab fatigue and distraction detection, tuned for precision so it stays switched on.
Zone intrusion, PPE compliance, counting and dwell, processed on site with no imagery leaving.
The right target is the smallest one that clears the decision threshold, not the largest one the budget allows.
| Target | Latency budget | When it is right |
|---|---|---|
| Microcontroller | 1–50 ms | Battery devices, always-on wake, anomaly detection on a sensor node |
| Embedded SoC with NPU | 5–50 ms | Vision at the machine, in-cab monitoring, multi-camera edge nodes |
| Edge server | 10–100 ms | Multi-stream analytics, heavier models, site-level aggregation |
| Custom ASIC accelerator | Sub-ms to ms | Volume products where power, cost or latency will not close otherwise |
| Cloud | 100 ms+ | Retraining, fleet analytics, anything without a hard real-time constraint |
Ask what the false-reject cost is on your line, and what it will be after six months of drift. If the answer is an accuracy percentage, they have not deployed one.
Feasibility studies, data strategy and annotation, model development, edge deployment and optimisation, computer vision engineering, sensor fusion, AI accelerator and silicon design, and MLOps for deployed device fleets.
Edge AI runs inference on or near the device, so decisions are not gated by network latency or availability and high-rate data such as video need not leave the site. Cloud provides compute for training and fleet analytics. Most production systems use both.
No. Dimensional gauging, presence checks and code reading are often better served by classical methods, which are deterministic, explainable and do not drift. Learned models earn their place where appearance varies beyond what a rule can express.
That is a successful outcome and usually saves a multiple of the study's cost. It normally comes with an alternative — a sensing change, a different measurement point, or a process change that removes the need for detection.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.