Ninety per cent of models
never leave the notebook.
Hire for the other ten.
Engineers who can frame the problem, build the feature pipeline, defend the evaluation, ship the thing behind a flag and watch it drift. On our payroll, inside your team, from $40 an hour.
Ask a candidate for one of these
This is the deliverable, not the notebook. It takes an afternoon to write and it is the single fastest way to find out whether someone has ever put a model in front of a business that depends on it.
Flags inbound insurance claims that warrant a manual review. It is a triage aid for an assessor, not an approval or denial decision.
Any use that reaches a customer without a human in between. Pricing. Anything protected-class adjacent. Written down so it is a decision rather than a drift.
Four years of settled claims, 1.2M rows, from two source systems. Twelve per cent were excluded because the outcome field was entered after the fact.
Recall at a fixed assessor workload — not accuracy. The business can review 300 claims a day, so the model is tuned to that budget, not to a leaderboard.
Weaker on commercial policies, which are 6% of volume and 40% of value. Those are routed to a human regardless of score.
Feature drift and score distribution tracked weekly. Ground truth arrives 60 days late, so performance is reported on a 60-day lag and everyone knows it.
The headline number is never the interesting one.
Two slices under 0.65. Both are named in the card, and both are routed to a human — which is a design decision, not a failure.
Four things that surprise first-time buyers
We say these in the first meeting because the alternative is saying them in month four, when they cost a great deal more.
If it is really an LLM problemMost of the job is not modelling
Feature pipelines, leakage hunting, label quality, and the serving path. The model itself is usually a fortnight of the six months.
Offline metrics are a hypothesis
The only honest evaluation is a shadow deployment against live traffic, followed by a controlled rollout. We plan for both from the start.
Training/serving skew is the classic killer
The feature computed one way in the notebook and another way in production. A feature store or a shared transformation library, not discipline alone.
Retraining is a product decision
How often, on what trigger, with what approval. Left undefined, it means never — and the model quietly decays for a year.
Platform-agnostic on purpose
Three bands, 8–14 days to start
Mid-level
3–5 yrs$40–50/hrBuilds and evaluates models against a defined problem, with the pipeline and serving path designed by someone else.
Senior
5–8 yrs$50–62/hrOwns the problem framing, the feature pipeline, the evaluation and the rollout. Writes the model card.
Lead
8+ yrs$62–85/hrSets the ML platform and the standards, and is the person who says no to the use-cases that should not be models.
Describe the decision, not the algorithm
Who makes it today, how often, and what it costs when it is wrong. That is enough for us to say whether it is a model, a rule, or neither.