What students have said about the courses

// challenge

The student worked at a software consultancy where engineers regularly lost time searching internal documentation that had grown across multiple wikis and repositories. They wanted to build a retrieval system that could answer questions against that corpus accurately enough to be trusted for day-to-day use.

// approach through the programme

Worked through the retrieval systems module in weeks 17–20, then designed the capstone around hybrid dense and sparse retrieval with a small open-weight model for answer generation. The system had to run for one month before the capstone defence and produce answers that a colleague unfamiliar with the system could evaluate.

// what the capstone produced

A running system with an evaluation harness that tracked answer quality over the month. The reviewer's notes identified one retrieval failure mode that became the main thread of the defence discussion. The repository was transferred and the system has been running in the student's team since August 2025.

// challenge

The student had a working text classification model in a notebook. A team member needed to call it from another service. Moving it out of the notebook and into something that survived concurrent requests was the problem they had not solved before the course.

// what the course covered

Packaging the model, containerising it, handling request batching, setting latency budgets and running a load test. The load test script from the reference repository was adapted for their specific model. Written feedback on the final deployment noted that the warm-start configuration was unnecessarily conservative and suggested the adjustment to fix it.

// outcome

Survived the load test. The student noted that the monitoring setup — specifically the quality drift detection covered in week three — was the section they expected to skip and ended up spending the most time on.

// starting point

The student had been using transformer-based models via an API for two years. They understood what the models produced but could not explain what happened when prompt wording changed the output in unexpected ways, or why certain failure modes were consistent.

// through the course

Wrote tokenisation from scratch in week one. By week three, the attention calculation was clear enough that the student could read the original "Attention Is All You Need" paper and follow the notation. That was the stated goal for week three, and it is unusual for it to happen on schedule — it did for this student.

// six weeks later

By the final session the cohort had a model that produced recognisable (if bad) text. More practically, the student noted they were reading the library source code instead of just the documentation when something behaved unexpectedly. That is the shift the course is designed to produce.