Small models in your stack, trained on your data, and a team that can run them.
Four services, each scoped to your case and enquired about rather than bought. We integrate Intura's models into your applications and agent workflows, train custom models on the context only your company has, build software around them, and train your teams to work with data, models and agents. Whatever we build runs on a CPU or an edge device, where your data already lives, and is handed over as something your team owns.
Each service is scoped to your case, but the engagement runs the same way, so you know what happens next at every step.
1
First call
A 30-minute call with an engineer, not a qualification script. You describe the problem; we say which service fits, or that none does.
2
Written proposal
Scope, what you provide, what you receive, where it runs and the timeline for your case, in writing before any work starts.
3
Build with reviews
Work in your environment or a copy of it, reviewed with your team at each step rather than revealed at the end.
4
Handover
Code, models, documentation and a runbook, walked through with the people who will own them. Support afterwards is optional and agreed separately.
Machine learning, SLM, or fine-tuned LLM
Which one fits depends on the decision, and getting that choice right is our problem rather than something you should have to settle before talking to us. Whichever it is, it comes back compressed: trimmed, quantised, benchmarked, and small enough to serve where you need it.
machine-learning
Classical machine learning
Classification, extraction, scoring and ranking. Small, fast, cheap enough to run on every event — and the right answer far more often than a language model is for a fixed risk taxonomy.
small-language-model
Small language models
Language understanding that still fits on a CPU and inside your perimeter. Tuned to your domain and your language, so it beats a general model at the one job you actually need.
large-language-model
Fine-tuned LLMs
For the tasks that genuinely need one: adapted to your policies, your case notes and your knowledge, with an evaluation set that says whether it actually improved.
What you take delivery of
Every engagement ends with artifacts you own and can operate without us. Nothing on this list is a licence you keep paying for.
Whatever we built, whether code, models, datasets or a curriculum, owned by you
Running in the environment you chose: your VPC, on-premise, air-gapped or on the device
Documentation and a runbook your team can follow without us
A handover session with the people who will operate it
A named engineer for the length of the engagement
Support terms agreed for how you run it, if you want them
Clients
Trusted by teams at
Private & edge deployment
Your infrastructure, ours, or no server at all
Because the models are small, where they run is a choice rather than a constraint. A 52 MB INT8 graph fits on a CPU box, a small GPU, an ARM board or inside a browser tab — so the private and edge options here are real options, and they get the most attention because they are the ones most vendors in this category cannot offer at all.
On-premise, your VPC, or air-gapped
Your own infrastructure
One Docker Compose file and one make target. It joins the network you already have, uses the redis you already run, and reads weights from a bucket you own — so a new checkpoint is a file swap rather than a rebuild, and no message, prompt or document crosses your boundary in either direction.
One command to deploy: weights, build, start, health check
Weights mounted read-only from a bucket you own, never baked into the image
Air-gapped is a supported configuration: no endpoint makes an outbound call
Scales across workers on a single box
Benchmarked on your hardware before handover, with the runbook
Buy once, run it yourself
Self-serve download
Pick a model in the marketplace, log in, and download it. You get every version from your library and run it with onnxruntime on whatever hardware you have.
Every version included with one purchase
Files from cloud storage, each with its SHA-256
Real example results on every model page, before you get it
Embedded, mobile, WASM
On-device & browser
The artifacts are standard ONNX. Load them with onnxruntime-web, on an ARM board, or inside a mobile app, and the check never touches a network at all.
INT8 graphs, 12–77 MB each
WASM and WebGPU execution providers
Offline by construction, not by configuration
Enquire
Tell us what you need
Describe the problem, the team and where it has to run. An engineer reads every enquiry and replies by email.
You will get an engineer on the first call, not a qualification script.
FAQ
Questions about Enterprise
▶▼Which service do I need?
Describe the problem on the first call and we will say. Many engagements combine two: a custom model and its integration, or a custom app and training for the team that will use it. Each service has its own page if you want to read ahead.
▶▼How is Enterprise work priced?
Each engagement is scoped on the first call and quoted in a written proposal. There is no checkout for Enterprise work, and nothing about it is a subscription.
▶▼Can the work happen inside our environment?
Yes. Integration, training data and deployment can all stay inside your VPC, on-premise or air-gapped when your policy requires it. Say so early; it is a normal requirement and we scope for it.
▶▼Do you need a GPU to train a custom model?
Not one you own. Training is our side of the engagement; the model you run afterwards is CPU-only, like everything else on this site. If your policy requires training to happen inside your own environment, we scope for that.
▶▼Can you build a model small enough to run on the edge?
That is the specialisation. Every model on this site was made small on purpose: the vocabulary trimmed to the languages served, layers cut, the graph quantised to INT8, and the model re-checked after each step. We do the same for a custom model, and tell you what each compression step cost.
