Research areas
An AI think tank that ships
Every capability here exists because a client needed it and we built it. Nothing on this page is a slide.
Research areas
Language and text
Pulling names, topics and meaning out of text at scale: entity extraction, topic modelling, summarisation, sentiment, contextual search and article recommendation.
Generative AI
Assistants grounded in your own documents, private models installed on your hardware, models tuned on your material, and image generation with fine-grained control.
Agents and tools
Agents that plan and act on real systems, reaching your databases and APIs through a standard interface any assistant can connect to. On-premise or cloud.
Computer vision
Understanding what is in a frame, matching it against what has been seen before, and making degraded footage legible again. Includes mask face recognition — identifying people whose faces are covered — which very few systems attempt at all.
Audio and speech
Transcription of calls and recordings, translation across languages, speech synthesis, voice identification, and voice interfaces onto systems that had none.
Analytics and model design
Model design and training, recommendation and ranking, and the algorithms underneath the products — the part that is genuinely research rather than integration.
Where we work
- News and publishing
- Broadcast and OTT
- Education
- Travel and hospitality
- Retail and food service
- Defence and government
- Enterprise
Privacy and deployment
Your archive never has to leave the building
Most AI vendors need your data to travel to them. For a newsroom protecting a source, a defence installation, a hospital or a bank, that is not a trade they are permitted to make. For everyone else it means handing over contracts, records and internal know-how, and trusting somebody else's retention policy with it. Neither is necessary.
A private AI box
A self-contained machine in your own facility. The models, your data and every inference run on it. There is no account to create and no vendor to trust, because nothing leaves the box.
Private by architecture, not by policy
A promise not to send your data anywhere can be undone by a configuration change. A system with nowhere to send it cannot be. We build the second kind.
Open models, deployed on your hardware
Strong open-weight models installed on your machines and yours to keep. No per-token billing, and nobody deprecating the model you built on.
Sized to the hardware you have
Models tuned to what is realistic for you, from a single GPU workstation to a rack. The enhancement toolkit runs on one GPU-equipped laptop.
Your archive stays your archive
For a publisher, the archive is the asset. Exposing it to an agent should not mean handing a copy of it to a third party, and here it does not.
Yours to audit and to keep
Your team gets the deployment, the documentation and the ability to inspect what the system does. Built to be handed over.
How an engagement runs
Six steps, in this order
Structured so the expensive decisions come after the cheap ones, and so what gets built is the thing that was actually needed.
Understand the goal
What the business is trying to change. Not which model to use — that question comes much later.
Map the data
The pipelines you have, the ones you did not know you had, and the ones that will need building.
Define the solution
A tailored proposal and an implementation roadmap, sized to what your organisation can actually absorb.
Prove it
A proof of concept against your real data, before anyone commits to an architecture.
Build
Production development, with your team close enough to the work to own it afterwards.
Test and hand over
Testing, deployment, documentation and handover. Systems you own rather than systems you rent.
Custom work
When the answer is not a product
Most of what is on this page ended up inside something we sell. Some of it did not, because the client needed something specific to them. We take that work too — a research partnership, a proof of concept against your data, or a system built and handed over.
Bring us the problem, not the technology
If it involves getting a model to do something useful with your own material, it is probably worth a conversation. If it is not, we will say so early.