What Participants Say About the Work

Challenge

A backend engineer with two years of ML exposure was producing models that performed well in notebooks and poorly in staging. The preprocessing steps were not documented, splits were done inconsistently and baselines were chosen by feel rather than by comparison.

What the cohort addressed

The nine-week curriculum gave explicit methods for documenting schema decisions, cleaning without destroying signal and choosing baselines through structured comparison. The peer review pairing meant a colleague was reading and commenting on the same decisions each week.

Outcome

By the end of the cohort, preprocessing code included written commentary on every non-trivial decision. Staging performance matched notebook performance for the first time. The participant described it as building the habits that should have been there from the beginning.

Challenge

A data analyst with prior experience in tabular data needed to build a visual inspection tool for a manufacturing context. She had no background in convolutional networks, no experience with object detection frameworks and no access to GPU hardware of her own.

What the specialisation provided

The fifteen-week structure moved from preprocessing foundations to detection and transfer learning at a pace that accommodated her starting point. GPU credits removed the infrastructure barrier. The third build — the deployed system — was the deliverable she needed to test the idea in practice.

Outcome

The final build was a detection system deployed to a low-cost single-board device, running inference within the latency budget required for the application. The portfolio review gave her a written document she could use in subsequent discussions with the team she was presenting the work to.

Challenge

A software engineer with three years of production backend experience had been asked to take ownership of a model that had been deployed informally by a research team. There was no versioning, no monitoring, no documented serving architecture and no clear understanding of when it was degrading.

What the programme covered

The nine-month structure addressed exactly the conditions he had inherited: reproducibility, versioning, pipeline orchestration, containerisation, drift monitoring and incident response. The five engineering-reviewed checkpoints tracked progress against the actual team project, not against a hypothetical.

Outcome

By the end of the programme, the model he had inherited was containerised, versioned and monitored with drift alerts that had fired twice and been handled. The architecture documentation produced in the writing workshop became the handover document for the next engineer who joined the team.