AI/ML, Analytics & Business Intelligence

The test a model has to pass

A model is worth building if someone changes a decision because of it. That is the entire test, and most analytics work fails it — not because the statistics are wrong but because no decision was ever attached.

Wanderhome’s model told us which house to buy and what to spend renovating it. That is a decision with six figures behind it, made differently because of a regression. Diagnostic.ly’s video grader decides whether a specimen collection was performed correctly, which determines whether a kit gets reshipped. Both change what happens next. Neither is a dashboard.

So the first conversation is always: what decision is currently being made badly, how often is it made, and what does being wrong cost? If there isn’t a good answer, we will tell you not to build the model.

Where we put AI, and where we refuse to

This is the part worth being precise about, because the market is about to spend several painful years learning it.

AI goes where being wrong is recoverable. Generating educational content. Watching a process and flagging a probable error for a human. Translation. Drafting from a structured template. Choosing which outreach to send next. Producing code, tests and documentation under senior review.

Deterministic rules go everywhere else. At Diagnostic.ly, clinical routing, ordering logic and result handling are explicit, auditable if-this-then-that logic that a compliance officer can read line by line. No AI diagnosis. No AI clinical recommendation. Not because the models can’t do it — because you cannot defend it, and the litigation is coming.

The same boundary applies outside healthcare. Credit decisions, pricing that touches protected classes, anything with a regulator or a plaintiff’s attorney attached: keep the decision logic legible.

What we actually build

Underwriting and pricing models. Regression and prevalence models over purchased and internal data, at the granularity where the decision is actually made — segment by geography, cohort by channel, SKU by market.

Scoring and prioritisation. Lead scoring, engagement monitoring, progressive profiling, coverage matching. Systems that rank a queue so a human spends their day on the right end of it.

Synthetic training data. When the labelled dataset you need does not exist — and for most real operational problems it does not — generating it is often more tractable than collecting it. We built a permutation engine for exactly this: procedure phase by error category, with Latin-hypercube sampling over the nuisance variables that wreck field performance, deterministically seeded so any subset regenerates identically.

Reporting people open. Which almost always means fewer numbers, attached to a decision, delivered where the decision is made.

AI readiness, honestly assessed

A large share of what we are asked for is an AI strategy, and a large share of what companies need first is data that is clean enough to model, infrastructure that can serve a model, and a governance position that will survive contact with their own legal department.

The four-to-six week assessment answers: what could you realistically deploy in the next two quarters, what would it be worth, what has to be true first, and what should you refuse to do.

Where we are strongest

Healthcare and regulated industries. Marketplaces and real-asset pricing. Ecommerce. And any business sitting on operational data it has never modelled — which, in our experience, is most of them.