Real-time ML that runs on the phone in your users' hands
We compress and tune models to run on real devices, fast, offline-ready and stable from flagships down to mid-range Android. The team behind Brickit and a mobile marking-verification scanner.
Discuss a pilot →Quick facts
- Business size
- Consumer apps and field tools with a camera or scanner, any size
- Timeline
- 6-10 weeks to a working on-device model, longer to harden across the device zoo
- Budget range
- Pilot from €40k. Cost scales with the platforms (iOS, Android) and the device range to support.
- Hardware
- Your users' phones, including mid-range Android, plus the cloud for the server side of a hybrid setup
- Data needed
- Labelled images for the task, plus access to real target devices for on-device testing
- Evolution
-
- Genesis
- Custom-built
- Product
- Commodity
No product gives you this. We assemble and train it around you.
What this scale means
Further right means more proven and cheaper. Further left means newer and riskier. Here is the test for each step.
- Commodity
- You could get the result yourself from a ready service, with almost no work. We rarely take these on.
- Product
- A vendor already sells this result turnkey, like shelf recognition from Trax or document reading from ABBYY. If one of them fits you, use it. You come to us when it does not: when it has to run on your own servers, cost less, or fit systems the product cannot reach.
- Custom-built
- Vendors sell only the parts. A tool like Tableau hands you charts, but the dashboards and metrics for your business still have to be built. That build is the work, and that is us.
- Genesis
- The approach exists but does not work reliably yet. You are betting on it maturing, so it costs more and carries more risk.
Expected outcomes
The Problem
A model that flies on a workstation can crawl on a phone. Real devices have limited RAM and CPU and a battery to protect. Speed ranges widely too, from this year’s flagship down to a three-year-old mid-range Android. On top of that, mobile users expect an answer in a second and they expect it to work in a basement with no signal. Most “AI in the app” demos send every frame to a server. That is slow, costly at scale, and useless offline.
What the Solution Does
We take a model that works and make it run on the device your users actually carry.
- Compress. Quantization, pruning and distillation shrink the model to a size and speed that fit a mobile budget.
- Convert. We target CoreML on iOS and NNAPI, MediaPipe or TF Lite on Android, handling the operator quirks of each path.
- Split. A hybrid design keeps the common, latency-sensitive path on the device and sends the rare heavy cases to a server. So the app stays responsive offline and gets sharper when a connection is available.
- Maintain. An update and monitoring pipeline ships new weights and watches on-device accuracy without waiting for an app-store release.
Where It Fits
This makes sense if you…
- Have a camera or scanner feature that needs to feel instant
- Need the feature to work offline or on a flaky connection
- Ship to a wide range of devices, including mid-range Android
- Care about the cost of running inference in the cloud at scale
This is probably not the right fit if you…
- Run a task that genuinely needs a data-center GPU per request
- Have a tiny, controlled device fleet where a server call is always fine
FAQ
Can you run our existing model on-device, or do we start over?
Usually we start from your trained model and compress and convert it. A retrain is only needed when the architecture does not fit a mobile budget at all.
iOS only, or Android too?
Both. The conversion and testing work is per platform, and the Android device zoo is the harder half because of how much the hardware varies.
What about battery and heat?
They are part of the budget. We profile on real devices and size the model so a continuous camera feature stays within a reasonable power draw.
Ready to Discuss?
Do you have a camera or recognition feature that needs to run on the phone, fast and offline? We will scope a pilot on your real target devices.
Related
Part of: Computer Vision ↗