Software, ML and AI engineering
Custom models, data pipelines and LLM applications, built so that every result can be validated and every run can be repeated.
Email us
What we build
Custom models
Trained on your data for a task no off-the-shelf model covers. Image, text or tabular, whichever the problem is in.
Data pipelines and MLOps
Ingestion, labelling support, versioning and retraining that keep working after the person who built them moves on.
LLM applications
Retrieval, tool use and structured output around a language model, with an evaluation set that decides what ships.
Validation and reproducibility
Held-out validation
Splits and metrics are agreed before training starts, so a result is a result and not a story told after the fact.
Reproducible training
Pinned data, pinned code, recorded runs. Any number in a report can be traced back to the run that produced it.
Documentation as you go
Approval documentation written after the project is documentation written twice. In regulated work we write it alongside the code.
Handover
You get the code, the weights and the training record. Nothing about the pipeline depends on us staying.
Showcase: treatment response from biopsy scans
A model on H&E-stained biopsy scans
We built the AI and ML pipeline that trains a model on H&E-stained cancer biopsy scans to predict a patient’s chance of responding to treatment. The long version is on the Work page.
What we need from you
A described task, a clear label definition, and access to a sample of the data. From that we can say whether the problem is tractable before anyone signs anything.
Talk to us about a model
Send the problem and a description of the data. We will say whether it is tractable.
Email us