FIT
It runs where it must
Sized and quantised for the silicon in the product, not for a rack. The constraint comes first and the model is designed around it, which is the opposite of the usual order.
Custom models
A generic model is a tourist. Yours should be a local.
Frontier models are extraordinary and completely unbothered by your problem. A small model trained on the twelve hundred things that can actually happen in your product will beat it on the only axis you care about — and fit on a chip you can afford.
Why bother
Small, fast, and embarrassingly specific
FLUENCY
It speaks your dialect
Room names, part numbers, alarm classes, the way your users actually phrase things at six in the morning. Domain vocabulary is where general models quietly fall over.
FINALITY
It is yours
No per-token meter, no deprecation notice, no provider changing the behaviour of your product on a Tuesday. The weights ship with the thing you sell.
Training a model is not the hard part. Knowing whether the new one is actually better than the old one — that is the hard part, and it is where most teams are guessing.
The pipeline
We built a factory, not a notebook
Custom models are only viable if producing the next one is routine. So the pipeline is the product we invested in first: data generation, training, evaluation, release — each stage repeatable, each candidate scored before anything ships.
01
Define the world
An explicit inventory of what the model must understand: the intents, the objects, the phrasings, the edge cases, the things it must refuse. Written down before a single epoch runs, because an undefined target cannot be measured.
02
Build the corpus
Real data where it exists, generated and augmented data where it does not, audited for balance and for the failure modes synthetic data loves to introduce. Coverage is a design decision, not an accident of what was lying around.
03
Train candidates
Several, deliberately — different base models, sizes and recipes — on our own GPUs. Candidates compete; the winner is chosen by the scoreboard rather than by whoever trained it.
04
Judge without mercy
Held-out and out-of-distribution evaluation, plus the cases the product genuinely cannot get wrong. A candidate that regresses anywhere that matters does not ship, however good the headline number looks.
05
Ship to the edge
Quantise, measure on the actual target — latency, memory, thermals, accuracy after quantisation — then release through the update channel with a way back if the field disagrees with the lab.
Worked example
Simone: the assistant that never leaves the house
Simone is the voice inside ALYT. Speech recognition tuned to how people actually talk to a house, intent understanding built around the rooms, devices and routines that exist in that home, and replies composed on hardware sitting in the building.
It is the honest version of the local-AI pitch: not a wrapper that phones a datacenter with the microphone open, but models we trained, evaluated and shipped — which is also why we can keep making it better for one specific job instead of waiting for someone else's release notes.
What we take on
Model work we say yes to
SPEECH
Voice in one domain
Wake word, recognition and intent for a product's own vocabulary, running locally.
VISION
Seeing one thing well
Detection and classification tuned to your scene, your lighting and your false alarms.
SIGNALS
Sensor fusion
Turning noisy multi-sensor streams into a decision a device can act on immediately.
REASONING
Small local reasoners
Structured, tool-calling models on a premise box for work that must not leave the network.
Your domain, your model
Bring us a job a general model keeps fumbling.
The vocabulary it mishears, the thing it classifies wrong at dusk, the reasoning it cannot do without the internet. That is the shape of a custom model brief.