OptiIris is a from-scratch iris recognition system designed for the edge. It delivers high-assurance biometric identity on the kind of hardware you actually deploy — phones, handheld scanners, kiosks — without a heavy server behind it. It enrols or verifies someone in about 26 ms on a laptop CPU — under 6 ms on a GPU — holds its accuracy across different cameras and populations, and stores every template in a scrambled, revocable form — so a leaked database exposes no one's iris. Because we built the recognition core ourselves, we can tune it for the accuracy and footprint your device demands.
Turns an iris image into a stable identity signature and matches it in about 25 ms — verify one-to-one, search one-to-many across a gallery, or screen for duplicates.
From a mobile phone to a handheld scanner to a large-scale server — one core scales down to the edge and up to the data centre, with no special hardware required.
Stays reliable across lighting, angle, and partial occlusion — and keeps performing on cameras and populations it has never seen, not just the one it was trained on.
Built-in quality and liveness checks resist presentation attacks, so a photo or replica doesn't pass as a live person.
Measured on a held-out test split, not a validation subset. The previously published 0.18% EER is a validation/subset figure; 1.41% is the honest held-out test number.
On the GB10 GPU every operation runs 1.8–5.4× faster than the published Apple M4 Pro CPU figures, and simple ops land at 4.7–5.9 ms — validating the ~2–10 ms GPU target. On a CPU-only edge device the same operations run ~34–56 ms.
Recognition runs on the device, and every stored template is scrambled and re-issuable — if a database leaks, the records are useless, can be revoked, and no one's iris can be reconstructed from them.
A from-scratch core means we can push accuracy and efficiency where a generic model can't, and adapt it to your use case.
Small and fast enough to embed, so identity works even with no connectivity.
Can be retrained and upgraded with easy rollback — it never changes silently underneath you.
Built for ARM and accelerated with MLX on Apple Silicon and CUDA on NVIDIA GPUs — tuned together for speed, accuracy, and lower power draw on whatever you deploy to.
All through one simple API, with each organisation's data kept fully separate.
Accurate enough to rival the best published results, fast enough for real-time use on everyday hardware, private enough that a data breach can't expose anyone's biometrics, and portable enough to run at the edge. Tested on standard iris datasets; independently reproducible.
Tell us what you're building. No sales pitch — just an honest conversation about the shortest path to shipping it.