ENNPRO Get in touch

Consulting practice

How we work on machine learning.

We start with a decision someone has to make, the evidence they have at the time, and the cost of getting it wrong. Those details determine what we build.

Related reading

The class that disappeared in the average

Model evaluation & repair

A single score can hide a recurring and costly mistake. We turn reported failures into a benchmark that the team can replay against the current model and its replacement.

Where we start

Bring a sample of disputed predictions and the reviewer’s reasons for changing them. Include the model version, the decision threshold and any inputs the operator saw.

What you take forward

A benchmark broken down by failure type, an error review and a release threshold tied to the cost of an incorrect decision.

Related reading

What does yesterday tell us about tomorrow?

Production ML engineering

A model ends up sharing a release schedule with data pipelines, product interfaces and people on call. We design the checks and recovery paths alongside the service.

Where we start

Bring the input contract, a representative historical feed, the serving deadline and the procedure the team currently uses when a prediction is missing.

What you take forward

A replay of the release, input freshness checks and a fallback that the operators responsible for the system can exercise.

The firm

Working with ENNPRO

See how we scope an engagement, review evidence with your team and prepare for handover.

Meet the practice