Specialized language models

The model comes to your data.

A frontier model is only available where your data isn't. Reaching it means sending regulated material out of the building — the exact thing your compliance programme exists to prevent. A model specialized to one domain is small enough to run beside the data instead, on hardware you already own.

Delivered

7 domains · each accepted before hand-over
Delivered

Edge and home

Residential AI

Turns spoken requests into device actions, compiles automations, and answers questions about live state — on a box in the building.

Delivered

Grid and energy

Utilities

Dispatch and curtailment commands, peak-shave and frequency-response logic, and grounded explanations of telemetry anomalies.

Delivered

Settlement

Financial services

Payment instructions into ISO 20022, legacy SWIFT translation, and explanations of returns and exceptions.

Delivered

Security

Cyber operations

Detection rules from described behaviour, ATT&CK technique mapping, CVE triage, and alert explanation from telemetry.

Delivered

Compliance

Governance and risk

Requirements mapped to NIST 800-53 controls, OSCAL output, control-family classification, and grounded gap explanations.

Delivered

Legal

Contracts

Clause classification and extraction, parties, obligations and dates, and risk explained against the clause it came from.

Delivered

Health

Clinical and payer

PHI removal, ICD-10-CM coding, entity extraction from notes, and answers grounded in the literature you supply.

Sectors we expect to fit next: Construction and field service · Manufacturing · Public sector records · Insurance.

What arrives

A model you hold, not an endpoint you call.

What you take delivery of is a file. It runs on your own machine, answers with no network available, and keeps working if we disappear. There is no key to revoke, no per-token meter, and no request leaving the building to be served.

Runs offline

No callback, no telemetry, no licence check. Air-gapped estates are the design case, not an exception.

Fits what you own

Sized for a workstation or a single node, so it can sit beside the data rather than in someone else’s region.

Yours to keep

The model is delivered to you. Its behaviour does not change underneath you because a vendor shipped an update.

Before hand-over

Four ways a specialized model can quietly disappoint you.

A model that is sharp on your domain but has forgotten how the world works, or has become easier to talk into something it should refuse, is not a good trade. Every model is checked against all four before it is handed over — and rechecked after it is compressed for your hardware.

Domain accuracy

It does the work it was specialized for, scored on cases it was never trained on.

General knowledge retained

Specializing it does not make it forget the everyday knowledge it arrived with.

Safety held

It is no easier to misuse than the model it started from. Checked after compression, not before.

Stable when compressed

It answers the same once shrunk to run on your hardware. Compression is where quiet regressions hide.

Where the work happens

Two paths. You pick, based on what you are allowed to move.

Path A · In your boundary

Training runs on your infrastructure

Your data never leaves. The whole cycle — preparation, training, acceptance — happens inside the perimeter you already control. This is the path for air-gapped estates, CUI, and anything a regulator has an opinion about.

Defense · Federal · Regulated health and finance

Path B · Hosted

We run the line and hand you the model

You supply the domain and the data under agreement, and receive the finished model to deploy yourself. Less to stand up, faster to a working specialist. Inference still runs wherever you put it.

Commercial teams without spare compute

Either way, the finished model runs wherever you decide to put it.

Bring us the job your cloud contract will not let you send.

Tell us the domain and how you would know the answer was right. If correctness is checkable, it is a candidate — and we will tell you plainly if it is not.