The AiT Stack / AI Model Development
AI that builds the AI your silicon was made to run.
ModelCat is an autonomous model builder for edge, embedded and IoT devices. Give it your data, your target chip and your limits on speed, memory and power — it architects, trains, optimizes and hardware-validates the model, then hands it back ready to deploy.
The bottleneck
Everyone has a model. Almost no one can ship it on time.
The perception model is the part that demos well. Getting it optimized, quantized and re-targeted onto the actual NPU, DSP or MCU — inside a real power and latency budget — is where months disappear. And the moment the silicon underneath changes, much of that work is done again from scratch.
- High risk the model misses its size, speed or power target
- Depends on scarce “unicorn” ML + embedded engineers
- Rebuilt by hand for every new chip and every new generation
- You set the constraints — accuracy, memory, power — and hold them
- Usable by developers, systems engineers and product owners
- Re-target to new silicon by regenerating, not rewriting
A model is portable. The work of fitting it to a chip is not — until you can generate that work instead of grinding it out by hand.
How it works · AI in the loop
Trained on thousands of architectures, so it can build the best-fit one for you.
ModelCat learned from thousands of model architectures, training methods and real outcomes. You describe the job; it searches that space, builds candidates, and measures them on real hardware — then keeps learning as new architectures land.
Upload your data
Bring labeled data, or start from an included open-source dataset.
Pick your target
Choose from a wide, growing range of supported chips — or ask for yours.
Set constraints
Size, speed and power — or let ModelCat decide what fits best.
Submit the job
ModelCat architects, trains and optimizes candidate models for you.
Review the set
Get a set of optimized models with precision measurements you can trust.
Explore & deploy
Drill into every attribute, pick one, export to TFLite or other formats.
↻ New chip, new sensor, new generation? Regenerate and re-validate — don’t rebuild.
What you get
An entire model team, without having to be an AI expert.
Check data quality
ModelCat inspects your dataset up front, so the model performs in deployment — not just on the bench.
Control the attributes
Set the execution speed, power draw and memory footprint your model must live within.
Built to order
Every model is architected from scratch to your spec — tested, and ready to deploy.
Anchored to the real world
A built-in hardware farm validates on physical silicon, so the numbers reflect real results.
Retarget across chips
Model Retargeting moves a proven model to new silicon fast — the same job, a different target.
Continually learning
New architectures and training methods are added over time, so you always have the best options.
Where ModelCat fits the stack
The stack builds the silicon. ModelCat turns it into shipped intelligence.
Our EDA 3.0 thesis runs from intent to yield — architecture, design, verification, manufacturing. ModelCat sits at the far end of that line: it takes finished silicon and makes it useful, generating the on-device model the chip was built to run. It pairs naturally with CraftifAI — one generates the model, the other generates the pipeline around it.
Generate the model from intent
Declare the data, target and constraints; get a hardware-optimized model back, validated on real silicon and ready to deploy.
Generate the pipeline that runs it
Capture, preprocessing, inference scheduling and post-processing, mapped to the target and re-targetable across compute platforms.
For a chipmaker, that combination is a design-win engine: it makes your silicon dramatically easier to adopt for on-device AI. It’s the same logic behind ModelCat’s work with leading semiconductor platforms —
Read the thinking
Why we’re watching on-device AI — and what ModelCat is publishing.
From Model to Motion: The Perception-Pipeline Bottleneck in Edge AI
The model is the easy 10%. The system between it and the silicon is where the schedule goes — and why deployment, not training, is the next decade’s moat.
Read on The Watchtower →One Spec, Every MCU: Rethinking IoT Firmware
Why work that doesn’t transfer across silicon is the real tax — and why “regenerate, don’t rewrite” is the pattern that beats it.
Read on The Watchtower →ModelCat launches eIQ® Model Creator for NXP devices
ModelCat’s autonomous builder, brought directly into NXP’s eIQ toolchain to turbocharge AI model development on NXP silicon.
Read on modelcat.ai ↗Supported chips & the hardware farm
See the growing list of targets ModelCat builds for, and the physical hardware farm it calibrates every model against.
Explore modelcat.ai ↗Let’s talk
On-device AI program stuck between the model and the silicon?
That’s the conversation. Thirty minutes on where you’re headed and which of our technologies gets you there — no slides required.
Book a 30-minute intro →