Applied machine learning lab

Machine learning that works in the field.

Applied machine learning for farming, offshore ground and marine science, built with domain experts and run in production.

How we work

  • Built with the expert

    The person who knows the domain sets the target: what the model has to get right, and how we'll know it has. For our wheat counter that was plants per 50 × 50 cm quadrat, chosen with farmers who run on-farm trials.

  • Into production, then better with use

    We host models on the InsightML Portal so they get used on real work. The aim is for the new data each one sees to improve its next version. That loop is at an early stage: our wheat counter's inputs have been captured from April 2026, and the first retrain on them is still to come. We also host trained models that other people have built (see the Portal).

  • Shown as it is

    Our model pages carry a model card and a build story: how the model was built, what it was measured on and where it falls short.

Featured · Geotechnical

Animation of a 3D block of the ground below the seabed at IJmuiden Ver Alpha and Beta. A vertical cut steps north through the site, exposing one slice at a time. Colours show the predicted soil-behaviour type: clay, silt-clay, silt-sand, sand, dense sand and very stiff. The block is mostly sand-coloured, with patches of silt-sand and clay.
IJmuiden Ver Alpha and Beta, Netherlands. Predicted Robertson soil-behaviour types in vertical slices, 200× vertical exaggeration. These are model predictions, not a geotechnical interpretation.

SyntheticCPT predicts what a cone penetration test would measure from 2D ultra-high-resolution seismic, so it estimates the ground between the real CPTs you have. Across three North Sea sites it has produced 454,030 synthetic CPTs, stacked into soil-behaviour models of the top few tens of metres. None of it needed a new survey.

Models

What we've built

All models →
Live
in production for clients
Beta
working, on limited release
Research
in development

Platform

The InsightML Portal

Sign in to the Portal →

Where our models run. Clients sign in to use our models in the browser or through an API, with usage counted per product. Each product carries its stage, from in development to production, so you can see how far along it is.

We also host trained models that other people have built. You keep the IP of your model. We convert it to a standard format, put it behind an API with key sign-in and usage counts, and add a basic demo page in the browser.

Research

Build stories and papers

  1. Paper · Conference abstractMedical imaging

    Diagnostic accuracy and clinical utility of AI-assisted transthoracic echocardiography for detecting left ventricular systolic dysfunction: a meta-analysis and diagnostic test accuracy

    Ryan Browne, Freddie Allen-Stevens, Niall Hill · European Heart Journal 47 (Supplement 1)

    A PROSPERO-registered systematic review and meta-analysis of AI-assisted echocardiography for detecting left ventricular systolic dysfunction. Across seven studies and 2,397 participants, pooled sensitivity was 0.82 and specificity 0.89.

  2. Build storyAgriculture

    Virtual Quadrat: a fixed square of ground in every field photo

    InsightML (Freddie Allen-Stevens)

    How a printed card with four markers lets farmers' phone photos be measured against the same nominal 0.25 m² of ground, what a season of 1,658 field photos taught us, and why an unreadable marker is still the main way it fails.