Blog
What we have learned shipping models and document pipelines into places where somebody signs the output, or where nobody can reach the device for a week. Written by the engineers who did the work, and signed.
Articles
What the EU is about to require of steel products, what the standards already require, and why reading a mill test certificate is harder than it looks.
Regulation · 6 minute read
Iron and steel is first in line for an ESPR delegated act, but none is adopted yet. What is settled, what is not, and what a processor should do meanwhile.
Read the articleEngineering · 6 minute read
OCR reads the characters and loses the rest: structure, units, heat scope. Why extraction needs a language model, validation rules and confidence scoring.
Read the articleStandards · 5 minute read
The difference is specific versus non-specific inspection, not price. A guide to the four EN 10204 document types and the situations each one belongs in.
Read the articleSeries
Notes from deployments where the camera is in a barn and the link drops for a day: what breaks between one site and forty, and what to decide before it does.
Field notes · 5 minute read
A pilot proves the model works at one site. Scaling to forty breaks on updates, connectivity, drift and ownership. What to agree before the pilot starts.
Read the articleField notes · 6 minute read
Latency is the usual argument for edge inference. On farms, solar arrays and in orbit the real reason is that there is no uplink. How to design for that.
Read the articleEngineering · 5 minute read
Most teams choose an edge device before optimising the model, then pay for that choice at every site. How to size hardware after optimisation, not before.
Read the articleSeries
Three posts from ELDER, a sheep lameness project on a working farm: the KPI we missed, the ground-truth problem that nearly sank the model, and how video became 200 bytes over satellite.
Field notes · 5 minute read
We missed one of eight KPIs on a livestock project. Here is the number, what it means at flock scale, and why accuracy is the wrong thing to ask about.
Read the articleField notes · 5 minute read
Our lameness classifier kept failing until we looked at the ground truth. Annotators were labelling two different conditions as one. The fix cost no compute.
Read the articleCase study · 5 minute read
Mountain grazing has no broadband. We ran the model on a Jetson in the barn and sent 200-byte events over a satellite link, a 99% cut against raw video.
Read the articleSeries
Audio-visual speech recognition on a neuromorphic accelerator: why fusion beats either channel alone in noise, what the hardware costs you in architecture, and a quantisation result we did not expect.
Case study · 5 minute read
An audio-only model at 3.7% word error rate returns 79.9% once the room is loud. Adding lip motion takes it to 16.6%. Measured numbers from a funded project.
Read the articleEngineering · 8 minute read
Measured with a power meter, not derived from operation counts. The protocol, our numbers, the caveats, and why you should benchmark your own model instead.
Read the articleEngineering · 4 minute read
Word error rate fell from 16.6% to 14.0% after quantisation, for the fused model only. The other three got worse. Here is the whole table and our best guess.
Read the articleNext step
If you are working through certificate automation, a Digital Product Passport programme, or an edge AI pilot that has stalled, describe it in a few lines. An engineer answers.
Free 45-minute call. Then a two-week data audit with a written go/no-go before you commit to anything. An engineer, not a sales rep, replies within one business day.