AI compiler / research in progress

Compile AI into tiny models.

A compiler layer that turns foundation-model representations into specialized models created from natural language and built to run anywhere. No task-specific data collection, no API costs, and complete user privacy.

Join the early list See how it works

TASK / NATURAL LANGUAGE

“Is this valve
open or closed?”

COMPILED MODEL

valve-state.onnx

8.7 MB · local

DescribeCompileDeploy

01 / THE CONSTRAINT

Foundation models do not solve every deployment problem.

Many AI use cases are constrained by network access, compute, privacy, or data.

TODAY

Train or call

Curate a dataset and train on expensive GPUs, or call a frontier model and pay for every image.

THE COMPILER LOOP

Compile once, run locally

Describe the task, compile a specialist model, and iterate without data curation or inference costs.

You own the model. Development stays fast, private, and local.

02 / THE COMPILER

Turn model capability into a reusable compiler target.

Small encoders can align their latent geometry with foundation models, allowing entirely new tasks to remain text-definable.

Once aligned, the shared representation becomes a stable interface between a general model and new specialist tasks.

Align once

Train a small model to speak the shared representation.

Define in text

Create new tasks without a task-specific labeled dataset.

Compile

Produce a specialist without another training cycle.

03 / THE FIRST PROOF

A visual compiler you can try live.

Align a 20 MB vision transformer with Gemma 4, then compile segmentation, detection, and classification models from natural language.

01DescribeDefine image classes in natural language.
02TestSee immediate behavior through a webcam.
03CompileCreate a small specialist without task-specific training.
04DeployDownload ONNX and run it locally.

MEASURE Teacher agreement Unseen-task generalization Local latency

04 / BUILT TO SHIP

The founder has built and shipped edge AI at scale.

PASSIO NUTRITION-AI

Real-time food recognition, entirely on device

Developed edge neural networks that recognized more than 4,000 food types at up to 30 FPS. Used by millions through partners including MyFitnessPal and Elevance Health.

4,000+: food types
30 FPS: on device
Millions: of users

MINDSEYE

Natural language to edge models

Compiled text-defined image classes into downloadable edge models.

~30 MB / 2,000 sign-ups

BYTEGPT

A minified language model

Built a compact language model designed for edge AI.

~85 MB / 1,000 Hugging Face downloads

QUANTUM BASE

Quantum authentication

Customized a contrastive neural network for quantum authentication.

Head of Software, Quantum Base PLC

James Kelly, PhD / Former CTO, Passio / Head of Software, Quantum Base PLC

05 / THE THESIS

Foundation models have already learned rich representations from vast amounts of data.

ijk compiles that learned structure into small models for specific tasks.

The resulting models solve real problems in environments constrained by compute, connectivity, privacy, or cost, where running a foundation model is impractical.

06 / EARLY ACCESS

See it when
it compiles.

Join the early list and I’ll let you know when the visual compiler is ready to try.