Route 1
Use Apple's built-in model
Good for
General text and quick features. Free, and already on the device.
Not for
A job specific to your product, a fixed budget, or behaviour that must not change with an OS update.
solutions
Put intelligenceinside your product.
Bring us one feature. We'll show it running on your device, with what it costs in power and memory.
by need
What teamscome to us for.
01
How we did it in [proof] →Work offline
The feature works with no connection, so there is no round trip and no failure state to design.
02
How we did it in [proof] →Keep data on the device
Nothing is sent, so there is nothing to store, secure or explain.
03
How we did it in [proof] →Fit a power or memory budget
You set the ceiling. The model is built to stay under it, and we show the measurement.
04
How we did it in [proof] →Run on the Watch
The smallest budget Apple makes. We have run a model on its Neural Engine.
where we fit
Three routes to intelligence on the device.Sometimes the first is enough.
Route 2
Drop in a ready-made model
Good for
Common jobs such as transcription or language detection.
Not for
Your own vocabulary, schema or tone, or a budget the stock model misses.
Route 3
Build one for your product
Made for the job
Start from one of our base models and tune it, with our tools or with us.
Your starting point
One feature, your own vocabulary, and a measured budget.
work with us
It starts withone feature.
The first engagement is fixed in length and price, and ends with something running on your own device.
01 / Fixed length · fixed price
[N] weeks · fixed price
On-device prototype
You bring
One feature of your app, and how it should behave.
We handle
The model, the conversion for Apple silicon, and a working build.
You receive
The prototype on your own device, and its measured power, memory and latency. If it won't run well, we say so.
02 / A couple of months
Product-specific model
The full build: model, AFKit integration, evaluation, and the interaction design around it.
03 / Ongoing
Keep it current
Your model re-measured and updated for every OS release and device generation.
proof
We did it toour own products first.
Uncle Rudy
A model that turns speech into structured tasks on the iPhone, with no cloud.
Measured cost: — mW · — MB · — MB · — ms
The Watch
A model running on the Apple Watch Neural Engine.
Measured cost: — mW · — MB · — MB · — ms
questions
Before youwrite to us.
- [Answer: today we build for Apple silicon only — iPhone, iPad, Mac and Apple Watch.]
- [Answer: who owns the tuned model and its weights at the end of an engagement, and on what terms.]
- [Answer: the prototype is measured on your device; if it won't run well, we say so and show the numbers.]
- [Answer: how our models and AFKit work alongside Apple's built-in model and frameworks.]
- [Answer: yes — AFStudio and AFKit are available to use on your own; we can help when you want it.]
Bring usone feature.
One paragraph is enough. A person replies within [N] working days.