Machine learning
The service we most often talk people out of. When it does fit, it’s because there’s a pattern in your data that a person would miss and a rule couldn’t capture.
Machine learning is worth it when three things are true at once: you have enough relevant history, the pattern is real but too subtle for a rule, and being right more often than chance is worth money. Take any one of those away and you’re better off with simpler software, and we’ll tell you so.
When it does fit, the work is mostly unglamorous — understanding what your data actually records, being honest about what it doesn’t, and building something whose errors you can live with. A model that’s right most of the time still needs a plan for the times it isn’t.
Fit
When this is the right job — and when it isn’t
We’d rather lose a project at this stage than sell you one you don’t need. The right-hand column is the honest half.
- You have years of relevant history and a decision that gets made repeatedly on it
- The pattern is real but nobody can write down the rule for it
- Being right more often than a coin toss is worth measurable money
- A person is currently doing the classification by eye and it doesn’t scale
- You have a few hundred rows. That’s a spreadsheet question, not a modelling one
- A handful of if-statements would get you most of the way. Simpler is better and cheaper to run
- You need to explain every decision to a regulator or a customer, and an interpretable rule would serve you better
- Nobody has agreed what “right” means yet. Without that there’s nothing to optimise toward
Scope
What’s included
An honest feasibility answer first
We look at your data before committing to a build. Sometimes that conversation ends with “don’t”, and that’s a good outcome for both of us.
A baseline to beat
We establish what the simplest possible approach achieves, so there’s a real bar. Plenty of models fail to beat a well-chosen rule.
Errors you can live with
Deciding which kind of mistake is more costly, and tuning for that rather than for a headline accuracy number.
Monitoring after launch
Models drift as the world changes. We build in the means to notice.
Related work
Where this shows up in the portfolio
Questions
Straight answers
How much data do I need?
It depends on the problem, but the honest signal is this: if you’re asking the question because you have a small spreadsheet, the answer is almost certainly that you don’t need a model. We’ll look at what you have and tell you plainly.
Isn’t this the same as AI agents?
No. Agents use existing language models to handle conversation. This is training something on your own data to make a specific prediction or classification. Different problems, different costs.