Custom detection models
Object detectors trained on your footage, with a focus on small, fast-moving targets and video from real conditions.
Computer vision & machine learning
We train machines to see, then fit them on a phone.
LOCKADOMAS builds custom vision models, from raw footage to on-device deployment, and tests each one on footage it never trained on. Our first product is an AI golf app. It's where we started, not where we stop.
- <1 ms
- NPU detection time per frame, field-tested on a Galaxy S26+
- 240 fps
- The capture rate our golf pipeline is built around
- 1,556
- Human-reviewed frames in our sealed golf benchmark, from 17 clips
- 0.02
- The most any metric can drop in quantization, or the model isn't delivered
What we build
The whole path from footage to a model on the device.
We started with one of the hardest things to see on a phone: a golf ball that can be about 5 pixels wide in the frame. The same system can be pointed at your footage.
Training data pipelines
A teacher model proposes labels, your current model's misses get mined, and a person reviews the frames that matter most.
Evaluation you can trust
Sealed benchmarks, leak checks, and a full scorecard of precision, recall, and mAP for every candidate model.
On-device deployment
8-bit quantization checked against the original model, then a field test on a real phone. Our work so far targets Snapdragon NPUs.
Also in the lab Audio event detection Multiple networks sharing one NPU AI agent tooling with permission gates
How we work
One system, eight stages, the same rules every run.
Capture
Collect real footage and set test clips aside before anything else.
Propose
A teacher model labels the footage so people don't start from nothing.
Mine
Run the current model and find what it missed.
Review
A person checks the selected frames and fixes the labels.
Train
Train several data mixes under the same rules, in the cloud.
Score
Compare every candidate with the current best model.
Quantize
Shrink the winner to 8-bit and measure what that cost.
Field test
Run it on a real phone before it's promoted.
Our one hard rule
No model grades its own exam.
- Test clips stay sealed Clips set aside to test a model never teach it, and clips that taught it can never test it.
- A person decides Models and scripts produce the scorecard. A person picks what moves forward and what gets promoted.
- Failures stay on the record When an experiment loses on untouched data, we log it and throw it out.
First product, in field testing
And yes, we're building a golf app.
Set your phone behind the ball and swing. It records at up to 240 frames per second, and neural networks on the phone find the ball and club head, with no upload and no server round-trip. We're not ready to show you everything yet.
Contact
Have footage? Let's see what a model can do with it.
We're taking on our first outside projects. Tell us what you need to detect, and we'll tell you whether we can help. Golfers, join the early access list here too.
joseph@lockadomas.com