Every project we take starts with the same two questions: what do you measure, and what decision hangs on it? Between those two points there is always a model. This page is about what separates a model that gets used from one that is eventually switched off: features that carry your domain's structure, predictions with calibrated confidence, explanations an expert can interrogate, and a deployment that learns from its mistakes.
We get there with a house method. Foundation models provide the features, fast models are what we ship, and conformal calibration sits on top. The details are on how we build.
Computer vision
Much of the scientific and industrial data that matters is a picture of something, whether it comes from a microscope, an inspection line, a field camera or a lab bench. We build detection, classification and measurement systems on modern vision backbones, including self-supervised models that learn your domain's structure from unlabelled archives before the first annotation is made. We specialise in the unglamorous cases that decide whether vision works in production, such as small objects, rare defects and imbalanced classes. Because the backbone yields features as well as answers, every system comes with attention maps and attributions that show what the model looked at, so an expert can tell a real detection from a lucky one.
Language & agents
Language models are the newest part of our toolbox, and we hold them to the oldest standard, which is measured behaviour. We build assistants grounded in your documents, structured extraction from decades of PDFs, generated reports that an expert signs off on, and agents that act with scoped permissions. All of it is built evaluation-first, so the question "how often is it wrong" has a number before anything ships.
Prediction & forecasting
This covers predicting quality from process parameters, properties from compositions and demand from history, as well as spotting anomalies in sensor streams. On tabular and time-series data we ship small models built on engineered and learned features. They train in minutes, are explainable by construction, and come wrapped in conformal intervals with a guaranteed error rate. When the model says 90%, it is right nine times out of ten, and that is a property we deliver and verify.
How we can help
Vision systems
Detection, measurement and anomaly finding on scientific and industrial imagery, with explanations attached.
LLMs & agents
Retrieval, report generation and tool-using agents, with the evaluation harness built before the demo.
AI Trust & Security
Red-teaming and audits for LLM systems you have already shipped, including ones we did not build.
Model validation & deployment
Is the model you already rely on calibrated, fairly evaluated and reading the physics rather than an artefact?
The method
How foundation models, domain adaptation, fast models and conformal calibration make our systems quick to retrain, explainable and clear about their uncertainty.
Deployment
Pipelines, Kubernetes and the active-learning loop that makes a deployed model improve instead of decay.
Workshops
Hands-on sessions for your team, tailored to your applications and, on request, to your own datasets.
Looking for the earlier service pages on statistics, classical machine learning and deep learning? They are still there, at the same URLs as always, marked as archived. We took them out of the menu because they describe what we sold in 2017 rather than what we do now.