Machine learning models
A model that scores 94% in a notebook and a model that earns its keep are different achievements. The distance between them is where this work lives.
The gap nobody budgets for
Training a model is the part that's well documented, well tooled, and genuinely quite fast.
Then it has to run somewhere. It needs data arriving in the same shape it was trained on. It needs to respond fast enough for whatever calls it. It needs someone to notice when the world shifts and its accuracy quietly decays. It needs to be retrainable by someone who isn't the person who built it.
That gap — deployment, pipelines, monitoring, retraining — is routinely underestimated and is where most ML projects stall. We treat it as the main body of the work, because it is.
What we build
- Classification
- Routing, categorisation, risk scoring, quality flags. Anything where the question is "which bucket does this belong in."
- Regression and forecasting
- Demand, pricing, capacity, lifetime value. Numbers you'll make decisions against.
- Ranking and matching
- Search relevance, candidate matching, prioritisation.
- Anomaly detection
- Fraud, faults, and outliers, in domains where the interesting events are rare by definition.
How we approach it
- We start with a baseline that isn't machine learning
- A simple rule, a lookup, an average. If that solves 80% of the problem, we'll tell you — and sometimes that's the whole project. The baseline also tells us whether the model is genuinely adding value or just adding complexity.
- We measure against the decision, not the metric
- Accuracy is a proxy. What matters is whether the prediction changes what someone does, and whether that change is worth more than the errors cost. We'll build the evaluation around your actual decision, including asymmetric costs — a missed fraud case and a false alarm are rarely worth the same.
- We show you the failures
- Every model gets a failure analysis: where it breaks, on what kind of input, and how bad the consequences are. An average accuracy figure conceals exactly the information you need.
- We build the pipeline alongside the model
- Training and inference share code paths, so the model sees production data in the same shape it saw training data. This single discipline prevents a large share of deployment failures.
What you get
Trained model with documented performance · Evaluation report including failure analysis · Deployment pipeline · Monitoring and alerting · Documented retraining process · Handover session
Before you ask.
It depends on the problem far more than on a rule of thumb. Some classification tasks work on a few hundred labelled examples; some forecasting needs years of history. We'll assess your data in the first week and give you an honest read before you commit.
Predictive Analytics
A forecast nobody acts on is a hobby. The prediction is the easy half; the hard half is what happens next.
Data & CloudData Engineering
Nobody has ever asked us for data engineering. They ask for the dashboard, the model, or the AI feature — and then we find out why it isn't working.
Tell us what you're trying to build.
A 30-minute call, no charge and no pitch deck. Describe the problem and we'll tell you how we'd approach it, roughly what it costs, and whether we're the right team for it. If we're not, we'll say so.
30 minutes · No charge · No deck