Custom model training
We fine-tune foundation models or train from scratch on your proprietary data. Classification, regression, sequence labelling, generative tasks. You own the weights and the inference code.
What we build
Four core capabilities. Each one is a standalone engagement or part of a larger system we design around your data.
Data pipeline engineering
Raw data in, clean features out. We build extraction, transformation, and loading pipelines that feed your models reliably. Airflow, dbt, Spark, or lighter tooling when the dataset is modest.
Inference API deployment
A trained model sitting in a notebook is useless. We containerise it, attach monitoring, set up autoscaling, and give your engineering team a documented REST endpoint. Average cold-start under 400 ms on GPU instances.
Ongoing model monitoring
Data drift kills accuracy quietly. We set up dashboards that track prediction distributions, input feature drift, and latency. When a retrain is needed, the pipeline triggers automatically or alerts your team.
How a typical project runs
No 90-page discovery document. We scope tightly, build fast, and iterate with your feedback every week.
Week 1: data audit and problem framing
We sit with your domain experts (usually over a video call, sometimes on-site) and look at the data you actually have. Not the data you wish you had. By Friday you get a one-page scope document: what the model will predict, what accuracy threshold matters, and what infrastructure we need.
Weeks 2–3: baseline model and pipeline
We stand up the data pipeline first, then train a simple baseline. This might be logistic regression or a small transformer, depending on the problem. The point is to have a number to beat before we invest GPU hours on bigger architectures.
Weeks 3–5: iteration and evaluation
Experiment tracking in MLflow. We try different feature sets, model sizes, and hyperparameters. You see results on a shared dashboard, not in a slide deck three weeks later. If the baseline was already good enough, we skip ahead.
Week 5–6: deployment and handover
The winning model goes into a Docker container, behind a load balancer, with Prometheus metrics and Grafana dashboards. We write a short runbook for your ops team. If you want us to keep monitoring, we offer a monthly retainer.
Common questions
Answers to the things clients ask us most often, before the first call.
How much does a typical engagement cost?
Most projects land between £18,000 and £65,000, depending on data complexity and deployment requirements. A straightforward classification model on structured tabular data sits at the lower end. Multi-modal pipelines with real-time inference push toward the upper range. We quote a fixed price after the data audit, not before.
Do we need to give you access to our production systems?
Not necessarily. We can work on anonymised exports in our own cloud environment. If you prefer us to deploy directly into your AWS, GCP, or Azure account, we do that too. We sign your NDA and DPA before touching any data.
What if our data is messy or incomplete?
It almost always is. Cleaning and enriching data is part of what we do. During the audit we flag gaps and suggest whether to fill them with synthetic data, external sources, or a simpler model that needs fewer features. We have never turned away a client purely because the data was rough.
Who owns the trained model?
You do. Weights, training scripts, evaluation notebooks, and deployment configs are all handed over at the end of the engagement. We retain no copies unless you ask us to for ongoing monitoring.
Can you work with our existing engineering team?
Yes, and we prefer it. Embedding with your engineers means the knowledge stays inside your organisation after we leave. We pair-programme, review pull requests, and document decisions in your wiki or Confluence instance.
Talk to us
Describe the problem you want AI to solve. We will reply within one working day with an honest assessment of whether we can help.
