AI Red logo AI Red
Contact
AI Red logo

Custom AI · Computer vision

Neural networks for problems off-the-shelf models can’t solve.

We build specialized vision models for fine-grained classification, on-device inference, and the data pipelines that make them work. Small, fast, and accurate — trained on your data, run on your hardware.

What we do

Built for the cases that break general-purpose AI.

Foundation models are great until the labels matter, the latency budget is tight, or the data is yours alone. That’s where we come in.

Fine-grained classification

When the difference between classes is sub-millimeter, easily occluded, or invisible without context. We design models that pay attention to exactly the right features.

On-device inference

ONNX, CoreML, and Unity Sentis models that run in real-time on iOS, Android, and edge hardware. No round-trip to a server, no cold starts, no per-request cost.

End-to-end pipelines

From capture protocol to labeling tools to nightly retrains. We build the whole system so your dataset keeps improving after we’re done.

Current work

Identifying RF connectors from a phone camera

A working example of the kind of problem we take on. Live, iterating, and instructive about how this category of model actually behaves in practice.

The problem

Ten classes that look almost the same.

SMA, 1.85mm, 2.4mm, 2.92mm, and 3.5mm RF coaxial connectors × male/female. Some pairs differ by 0.6mm of physical size — invisible without a reference scale. Photos vary by distance, lighting, finger occlusion, and the angle of a hex nut. The model has to run on commodity phone hardware and abstain rather than guess wrong.

Where we are right now

100%
accuracy on the most recent 410 phone uploads
0
confidently-wrong outputs (the metric we care about)
~420ms
per inference on commodity CPU

Have a visual problem off-the-shelf AI can’t handle?

Tell us what you’re working on. We’ll tell you if it’s a fit and what it would take.

or email chris@aired.com directly.