Solutions
AI products built for the constraints you actually operate under.
Purpose-built applications for a specific job — document understanding, speech capture, agentic reporting — including on-premise inference for environments where sending data to a third party is not an option.
The general-purpose assistant is not the product. It is the thing people try first.
Every organisation now has a chat assistant somewhere. Adoption is usually broad and shallow: people try it, find it useful for drafting, and return to their existing workflow for anything consequential. The reason is rarely model quality. It is that the assistant sits outside the system where the work happens and has no grounding in the organisation’s own data.
The products that stick are narrow. They do one job inside an existing workflow, with the relevant context already loaded, and they produce an artefact somebody needed anyway — a completed note, a formatted report, a structured record.
A second constraint shapes what is buildable: in healthcare, legal and defence contexts, the data frequently cannot leave the premises at all. That is a solvable architectural problem, not a reason to abandon the use case.
What we build
What we build in this space
Document understanding
Ingest, extract and analyse the document estate that governance and operations depend on — design records, audit reports, contracts, test results — with every extracted claim tied back to the page it came from.
Speech capture and structuring
Recording through transcription to a structured, domain-appropriate record. Built as a mobile-first application where the work happens away from a desk, and shaped by the vocabulary of the specific profession using it.
On-premise inference
Speech and language models running on hardware you control, reachable over a private network. Where data residency is a hard requirement, this is the difference between a product and a policy exception nobody will sign.
Agentic report generation
A user describes the report they need; the system inspects the real data through a constrained query interface, emits a structured specification, and a deterministic renderer produces it. The model never writes executable code, and the output is always on-brand.
Retrieval that cites
Retrieval built directly rather than through a heavyweight framework, tuned to the corpus, and returning the passage behind every claim so the reader can check it in one click.
Evaluation harnesses
Regression suites over real examples with a graded rubric, so a prompt or model change is a measured decision. Without this, every model upgrade is an act of faith.
What changes
What this changes
The output is the artefact, not advice about it
A finished note, a rendered report, a populated record. Products that stop one step short of the deliverable get abandoned, because the last step was the tedious one.
Data residency stops blocking the use case
When inference runs on hardware you control, the review that would have killed the project becomes a straightforward architecture conversation.
Changes become measurable
With an evaluation harness in place, swapping a model or revising a prompt produces a number rather than an argument, and regressions are caught before users find them.
Spend is attributable
Every model call is metered and tagged to the feature that caused it, so the unit economics of each capability are known rather than inferred from a monthly total.
Describe the workflow, not the model.
The useful conversation starts with the job somebody is doing by hand and the constraints around the data it touches.