We train visual inspection models from photos of good parts only and deploy them on a €80 board beside your line — delivering a PASS or FAIL verdict in under one second, with no network call, ever.
All numbers from benchmarks run in Munich, October 2026. Reproducible. Not marketing language.
Four forces push inference onto the device. Wherever any one applies, the standard cloud-AI playbook stops working — and that is where Nuatara operates.
A reject arm fires in milliseconds. A cloud round-trip is already too late. 120ms on-device beats 300–2000ms cloud every time.
EU AI Act, NIS2, data-localisation rules. Production images cannot leave the facility. Nuatara makes zero network calls at inference — by design.
Field installations, moving vehicles, remote sites. The network link is unreliable or absent. The system must run regardless.
Streaming video to a cloud inference server is billed per image, indefinitely. €80 hardware with a trained model on it costs nothing per inference, forever.
Train models. Have never touched a sensor driver, constrained silicon, or an ARM deployment.
Excellent hardware. No model training, no quantisation, no deployment pipeline.
Built for customers with in-house ML engineers. Priced and paced for them too.
Own the mechanics. Subcontract or skip the AI layer entirely.
The customer is left to integrate four vendors. That is where the months go.
Three steps from standing start to a working inspection system at your line. One team, one week, one measurable result.
Capture 50–200 images of good parts only. No defective examples needed. A phone camera works for a first test.
We train a PatchCore anomaly model on your good-part images. No defect labels. Training takes 15–60 minutes on a laptop.
Model deployed on a €80–250 ARM board. Run it on your parts. See a PASS/FAIL verdict with heatmap showing exactly where the defect is.
Fixed-price pilot per deployment. No open-ended retainer.
Five layers, one codebase. Swapping the backend is a config change. Swapping the source is what lets the same core handle vibration or thermal sensing later without a rewrite.
Every offering reuses the same pipeline — fit the model to the silicon, make it run, keep it honest. What changes is the sensor, the model, and the buyer.
All results from benchmarks run on Khadas VIM3 and Edge2 Pro boards. MVTec AD is the standard industrial inspection benchmark. Every number here is reproducible.
Nuatara was founded in Munich in October 2026 — on Dussehra, the Indian festival that marks the victory of truth over deception. We build technology that brings truth to the factory floor.
The founding insight: the hardest part of industrial inspection is not the AI model — it is removing every barrier that prevents a factory from trying. No defect labelling. No cloud. No six-month project. No €50k budget. First result in one week.
The October 2026 PoC confirms the hypothesis: PatchCore on a €80 ARM board, trained on 50 good-part photos, achieves AUROC 1.0 on two of three categories tested. The INT8 NPU path on the €250 Edge2 Pro brings total latency to 120ms.
"विजयादशमी — the day good conquers evil. We chose this day deliberately. Edge AI for industrial inspection is, at its core, about bringing truth and light to the production process."
Zero network calls at inference — a hard requirement, not a default that can be toggled off.
Running on a €80 board is not a compromise. Extreme hardware limits produce simpler, more robust systems.
AUROC 1.0, 120ms, 50 photos, one week. Numbers, not marketing language. Reproducible.
Early-stage. Focused. Looking for manufacturing design partners across DACH and Europe.
We are looking for design partners in manufacturing, energy, and logistics across Europe. One line, one detection task, one measurable result. One afternoon.
Write directly: hello@nuatara.ai