Machine learning for work where being wrong is expensive.

Methods

• Offline and batch RL from historical logs
• Off-policy evaluation with confidence bounds
• Contextual bandits and Bayesian optimisation where full RL is overkill
• Simulator and reward design, including reward hacking review
• Constrained and safe policies for regulated settings

In practice

Few organisations can let an agent act freely on a live system, so we start from the logs you already hold: feasibility, then off-policy evaluation, then shadow deployment against the incumbent policy. Where a simpler method matches an RL policy’s performance, we build the simpler method.

Methods

• EHR extraction, cohort definition and computable phenotypes
• Mapping to OMOP CDM and FHIR
• De-identification and information governance support
• Prognostic and diagnostic models reported to TRIPOD+AI
• Calibration, subgroup performance and fairness auditing
• Survival analysis and competing-risk models

In practice

The modelling is rarely the bottleneck. Producing a defensible cohort from a live clinical system, and agreeing what counts as an outcome, is where the time goes. We plan for that from the start and build to the reporting standard — the difference between an internal result and one that survives external review.

Methods

• Structure–property and composition–property modelling
• Design of experiments and active learning to cut lab cycles
• Image analysis for micrographs, porosity and scaffold morphology
• Surrogate models for expensive simulation
• Spectral and thermal data pipelines
• Degradation and shelf-life modelling from small samples

In practice

Materials datasets are small, expensive and unbalanced, which rules out most standard approaches. The gains come from designing the next batch of experiments well, not from fitting a larger model to forty samples. We work alongside the people running characterisation, because the measurement protocol is part of the model.

Methods

• Data audits: what you have, what is trustworthy, what is missing
• Pipelines, warehousing and reproducible analysis environments
• Causal inference and experiment design
• Forecasting and demand modelling
• Model monitoring, drift detection and retraining policy
• Analysis and reporting your team can run without us

In practice

A significant share of work that arrives labelled as an AI project is a measurement or data-quality problem. We are content for an engagement to end at that finding, with a fixed pipeline and a far cheaper answer than the one you budgeted for.

Methods

• Retrieval over your own documents, with citation back to source
• Evaluation sets built on your material before any model is chosen
• Structured extraction from clinical notes, lab reports and technical literature
• Fine-tuning and adapters where prompting has genuinely run out
• Guardrails, abstention and human-in-the-loop review points
• Cost, latency and vendor lock-in modelled before you commit

In practice

The demonstration always works. The real questions are performance on the ten-thousandth document, and whether anyone downstream would notice a fluent wrong answer. We build the evaluation set first, from your own material, and agree the acceptable failure rate before anything is built — including, where warranted, concluding that a language model is the wrong tool.

Methods

• Intrinsically interpretable models where accuracy allows
• Feature attribution and partial dependence, with their limits stated plainly
• Counterfactuals: what would have had to differ for another answer
• Saliency and attention maps for imaging, checked against expert judgement
• Calibrated uncertainty, and knowing when to abstain
• Explanation interfaces designed and tested with the people who use them

In practice

An explanation that satisfies a data scientist often fails the clinician, engineer or auditor who has to rely on it. We treat the explanation as part of the product and test it with its real audience. Where a simpler model is nearly as accurate and far more legible, we argue for the simpler model.

Methods

• EU AI Act risk classification and conformity readiness
• ISO/IEC 42001 management system design and gap analysis
• NIST AI Risk Management Framework mapping
• Model cards, dataset datasheets, intended-use statements, audit trails
• Data protection impact assessments and bias evaluation
• Post-market monitoring and incident response planning
• Software as a Medical Device classification and technical file support
• Independent review of systems you already run or bought in

In practice

Governance is inexpensive when designed in and costly when reconstructed before an audit. Most artefacts an assessor asks for are produced by a well-run project anyway, provided someone decided at the start to keep them. We also review systems built by others. We are not a law firm, and we say so when a question needs one.

Methods

• APIs and services around models, with versioning and audit trails
• Internal tools, research portals and annotation interfaces
• Public websites and product front ends
• Containerisation, CI/CD and cloud or on-premise deployment
• Accessibility and performance as delivery requirements
• Handover documentation written for your engineers

In practice

We build the surrounding software rather than handing over a notebook. That includes the unglamorous parts: logging, permissions, backups, and a runbook for the day something breaks while we are not there.