Custom models, judged against your economics.

A model is only worth building when a decision is repeated often enough, with a measurable outcome, that being right slightly more often is worth real money. Most of our work is establishing whether that is true before writing a training loop.

01 / The problem

Where this usually goes wrong.

The failure mode is rarely a bad algorithm. It is a model optimised for a metric that does not correspond to anything the business cares about — accuracy on a balanced test set, when the real population is 2% positive and the two error types have wildly different costs.

The second failure mode is subtler and more common: the model is trained on information that would not have been available at the moment of prediction. It scores brilliantly in testing and collapses in production, and by the time anyone notices, it has been making decisions for a month.

02 / What we build

The parts of the system.

A framed decision: Which decision this model informs, who or what acts on it, how often, and what the action costs when it is wrong in each direction. Written down before anything is trained.
Feature pipelines with a time discipline: Features computed identically in training and serving, with an explicit cutoff so that nothing from the future leaks into a prediction about the past. This is the most common cause of models that fail on contact with reality.
The simplest model that clears the bar: Frequently gradient-boosted trees on tabular data. We move to something heavier when it earns its keep on your evaluation set, not because it is more interesting to build.
Serving, monitoring and a retraining path: The model behind an API with defined latency and cost, monitored for input drift and outcome degradation, with a documented, rehearsed procedure for retraining and rolling back.

03 / How it is measured

What we agree to be judged on.

Set before the build starts, against a measured baseline, and reported honestly afterwards — including where the numbers are disappointing.

Against the incumbent: Whatever makes this decision today — a rule, a threshold, a person — is the baseline. Beating a random guess is not an achievement.
On a temporal split: Trained on the past, tested on the future. Random splits flatter almost every model built on data that has any time structure, which is most business data.
In business units: Rupees recovered, hours saved, cases correctly escalated. Precision and recall are how we tune it; they are not how we justify it.
Across segments: Performance broken out by region, customer type, product and channel. Aggregate accuracy routinely conceals a segment where the model is worse than useless.

04 / Honest limits

What this will not do.

Machine learning extrapolates from history. Where the future genuinely differs from the training period — a new product, a regulatory change, a competitor entering — it will be confidently wrong, and it will not warn you.

It also cannot manufacture signal that is not present. If the outcome you want to predict is largely driven by information you do not record, no amount of modelling will recover it, and the honest deliverable is a recommendation about what to start capturing.

05 / When it applies

You probably need this if:

A high-volume decision is currently made by a static rule or a person's judgement
You have several years of outcome data for that decision
Being right slightly more often is worth a meaningful amount
A previous model was built but never made it into production

Recognising two or more of these is a reasonable trigger for a diagnostic. Recognising none of them is a reasonable trigger for not spending money here yet.

06 / Related

Often scoped together.

These capabilities share data, infrastructure or evaluation approach with machine learning, and are frequently part of the same engagement.

Next step

Is machine learning the right instrument for your problem?

The diagnostic exists to answer exactly that, including the possibility that the answer is no. It is time-boxed, fixed-fee, and the written assessment is yours.

Request a diagnostic All capabilities