AI for UK Healthcare, with Clinicians in Charge

It depends on intended use. Software that handles administration — booking, routing, drafting letters a clinician edits — is generally treated differently from software that diagnoses, predicts risk or recommends treatment, but features such as summarisation can move closer to the line depending on how they are described and used. The MHRA regulates medical devices in the UK; we raise the question in discovery so you can take regulatory advice before the design is fixed.

We build an evaluation set from a de-identified sample of your own letters, notes or images, labelled by your clinicians or admin staff. The system is measured on it field by field and by document type, with particular attention to omissions — a missed allergy matters more than a misspelt name. Results go to your clinical safety lead before anyone relies on the output.

Summaries are generated from the source notes only, and every line links back to where it came from. We test specifically for omissions of high-risk items — allergies, medications, safeguarding flags — and show those items in fixed sections rather than leaving them to the model’s judgement. The clinician still reads the source for anything that affects care.

Wherever the task allows. Names, NHS numbers, addresses and dates can be removed or replaced before text reaches a model and re-attached afterwards inside your system. Some jobs, such as matching a letter to the right patient, need identifiers, so we keep that step inside your environment and document the data flow for your DPIA.

Yes. Several hosted model providers offer UK or European processing options, which we confirm for your chosen provider during design, and open models such as Llama or Mistral can run inside your own cloud account. The choice depends on your IG requirements, accuracy on your material and running cost.

Yes, as a draft. Speech-to-text produces a transcript, and a language model turns it into a letter or note in your template; the clinician edits and signs it before anything enters the record or reaches a patient. Patients should know recording is happening, and your IG and clinical safety leads decide how consent and retention work.

A dental health platform that uses computer vision to analyse dental X-rays and intraoral images and flag possible conditions for dentists to review, built with Python and TensorFlow on AWS. It follows the same pattern we apply elsewhere in healthcare: the model highlights, the clinician decides.

We are based in London, with onshore and offshore engineers in the UK and India. Most development runs on de-identified or synthetic data, so engineers rarely need identifiable records. Where live access is unavoidable, it is limited to named people you approve, and logged. If your policies rule out access from outside the UK, tell us early and we will state plainly what we can and cannot provide.

A discovery sprint to fix the intended use, check the data and plan the evaluation starts from £2,000. The pilot is then quoted at a fixed price, with evaluation and assurance documentation shown as separate lines so the effort is visible. Model and API fees are paid to the provider, and we estimate them up front.