Intelligence, built for the product. Tuned for the silicon.
We make models, tools and kernels for Apple silicon, and publish what each one costs in power and memory.
How intelligence got here
1956
AI is named.
1966
The first chatbot.
1997
A machine wins at chess.
2012
Deep learning, on GPUs.
2022
Large models, in the cloud.
2026
Small models, on your device.
this is where Scissor Works comes in
Four claims, one stack.
01
A product is intelligent when the intelligence is part of it.
Borrowed from a server, it waits on a network and sends data away.
02
So it should live where the product lives: on the device.
Offline, instant, free per use, and it keeps what it sees.
03
A device has limits: memory, battery, one kind of silicon.
Built for one product, it only has to do that product’s work.
04
So we build from the silicon up, and measure every layer.
Kernels decide power. Models, memory. Tools, what you ship. Apps, what a person feels.
THE RESULT
A product that is intelligent by itself,with the cost printed on the label.
One material, cut four ways.
The shiny thing is the intelligence. We cut it to the shape of the job.
Kernels
closest to the silicon
power · — mW
Models
the smallest that does the job
memory · — MB
Tools
convert, profile, ship
disk · — MB
Apps
private, offline, instant
latency · — ms
Closest to the silicon.
Borrowed from a server, the product waits on a network, pays for every use, and sends its users' data away.
— ours - - general model (shape is a placeholder)
Bring us the problem.
We bring the model.
Most teams come to us with one of these. Each is a feature we can prototype on your own device, at a fixed price.
Book a 30-min call
“Our AI feature needs the internet.”
We move it onto the device. Works offline, answers at once.
users →
“Our cloud bill grows with every user.”
On the device it costs nothing per use.
“We can’t send user data to a server.”
It stays on the phone. Nothing leaves.
“Apple’s built-in model isn’t enough.”
We tune one for your job, your vocabulary, your budget.
Notes from the workbench.
Every model, measured on a named device.
[One-line summary of the piece: what was measured, on which chip, and what it cost.]
Read the piece →
[Latest interaction piece title]
[Date] · [N] min
[Latest research piece title]
[Date] · [N] min
Three routes. Sometimes the first is enough.
Apple’s built-in model
General text and quick features. Free, already on the device.
You might already be done.
A ready-made model
Common jobs like transcription or language detection.
Good parts exist. Use them.
One built for your product
Your job, your vocabulary, your power and memory budget.
This is where we come in →
If route 1 or 2 is enough, we’ll tell you on the first call.
Work with us
It starts withone feature.
Bring us one feature of your app. We show it running on your own device, with what it costs. If it won’t run well, we say so.
01
On-device prototype
[N] weeks · fixed price. A working build on your device, with its power, memory and latency.
02
Product-specific model
A couple of months: model, AFKit, evaluation and the interaction design.
03
Keep it current
Ongoing: re-measured for every OS release, from the Mac to the Watch.
