Overview
You build one model: kestrel-late-risk at Kestrel, a UK parcel-delivery company with a Leicester hub, nine depots and about 210,000 parcels a day. The model predicts the chance that a parcel misses its promised day, so Kestrel can warn customers at 06:00 and send its few spare vans where they are needed. A prototype from last spring scored well, but nobody could rerun it, it trained on a column known only after delivery, and it has never seen a parcel from Doncaster, the depot that opened in August. You carry the real model from raw depot scans to a gated, benchmarked, monitored service, in one repository, with a decision record for every choice that needed one. Kestrel is fictional and its data is simulated; every lesson also opens with a real, cited case.
What you will learn
- 93 written parts, yours for good
- 58h 31m of reading, measured not estimated
- Written for the intermediate level
- Every future revision included
Curriculum
17 sections · 93 parts · 58h 31m3h 25m free to read · 55h 6m with the course
01Course guide6 parts · 71 min
- AI / ML Engineer: The Complete CoursePreview5 min
- How This Course WorksPreview7 min
- Setup: Colab, a GPU if You Want One, Model Providers, Libraries and PricesPreview12 min
- The Project: KestrelPreview13 min
- Shared Kestrel DataPreview27 min
- Run Locally: the Repository, Ollama and the Parts Colab Cannot RunPreview7 min
02Module 1: Mathematical Foundations6 parts · 134 min
- Module 1: Mathematical FoundationsPreview1 min
- Lesson 1: Parcels as vectors: the linear algebra of a modelPreview28 min
- Lesson 2: Walking downhill: gradients and how models learnPreview29 min
- Lesson 3: How sure are you? Probability for model buildersPreview29 min
- Lab: The Kestrel repository and a first baselinePreview32 min
- Incident: A squared probability went before a juryPreview15 min
03Module 2: Software and Systems Engineering5 parts · 106 min
- Module 2: Software and Systems Engineering1 min
- Lesson 1: From a notebook to a tested Python package28 min
- Lesson 2: The systems around the code29 min
- Lab: Kestrel as an installable, tested package40 min
- Incident: A nightly install pulled the wrong package8 min
04Module 3: Classical Machine Learning6 parts · 159 min
- Module 3: Classical Machine Learning1 min
- Lesson 1: Six core models on one parcel table27 min
- Lesson 2: Neighbours, groups, maps, forecasts and oddities29 min
- Lesson 3: Honest models: leaks, baselines, costs and calibration29 min
- Lab: Model v1, honest baselines and ADR-00158 min
- Incident: The model learned that asthma lowers risk15 min
05Module 4: Deep Learning7 parts · 211 min
- Module 4: Deep Learning1 min
- Lesson 1: How a network learns, from gradients to normalisation32 min
- Lesson 2: Keeping training stable, and debugging it when not27 min
- Lesson 3: Architectures: matching the network to the data33 min
- Lesson 4: Frameworks: PyTorch, Lightning and a run you can repeat31 min
- Lab: A network against the trees, and ADR-00272 min
- Incident: A frozen layer that was not frozen15 min
06Module 5: Data Engineering for ML7 parts · 209 min
- Module 5: Data Engineering for ML1 min
- Lesson 1: Pipelines that can run twice: batch, stream and backfill26 min
- Lesson 2: Where the data lives: columns, partitions and tables24 min
- Lesson 3: Data you can trust: contracts, versions and labels30 min
- Lesson 4: Features that mean the same in training and serving25 min
- Lab: The daily data path, with a contract88 min
- Incident: A computed input was wrong for three weeks15 min
07Module 6: Experimentation5 parts · 173 min
- Module 6: Experimentation1 min
- Lesson 1: Tracked experiments and honest comparisons28 min
- Lesson 2: Discipline: one change, one locked test, one honest log27 min
- Lab: Model v2, found and tested honestly103 min
- Incident: 62 models, none fit for clinical use14 min
08Module 7: Training at Scale6 parts · 189 min
- Module 7: Training at Scale1 min
- Lesson 1: What a GPU holds, how fast it feeds, what it costs21 min
- Lesson 2: Many machines, one model, and surviving a crash27 min
- Lesson 3: Where training time goes, and how to win it back22 min
- Lab: What Kestrel's training costs, and ADR-004110 min
- Incident: A training run that hung without failing8 min
09Module 8: Model Optimisation6 parts · 219 min
- Module 8: Model Optimisation1 min
- Lesson 1: Making models smaller: quantise, prune, distil29 min
- Lesson 2: Making models faster: ONNX, batching, caching, compiling28 min
- Lesson 3: Measuring: percentiles, cold starts and cost per thousand26 min
- Lab: Model v2 in ONNX, benchmarked121 min
- Incident: The benchmark said quantisation was fine14 min
10Module 9: Serving and Inference5 parts · 202 min
- Module 9: Serving and Inference1 min
- Lesson 1: Batch, online and streaming: three ways to score26 min
- Lesson 2: A scoring service that bends instead of breaking28 min
- Lab: The scoring service, batch and online132 min
- Incident: One shared filter stopped every model15 min
11Module 10: MLOps6 parts · 256 min
- Module 10: MLOps1 min
- Lesson 1: A pipeline, a registry and gates that say no30 min
- Lesson 2: Shipping a model without betting the warnings on it29 min
- Lesson 3: Retraining on its own, inside a budget27 min
- Lab: Continuous training, gated and replayed155 min
- Incident: A model update shipped a broken setting14 min
12Module 11: Monitoring and Reliability5 parts · 242 min
- Module 11: Monitoring and Reliability1 min
- Lesson 1: What to watch when the model is live27 min
- Lesson 2: Running a live model: alerts, dashboards, incidents29 min
- Lab: September replayed through a monitoring job170 min
- Incident: Bad training data, found through revenue15 min
13Module 12: Foundation Models6 parts · 307 min
- Module 12: Foundation Models1 min
- Lesson 1: Adapting a small open model to Kestrel's messages34 min
- Lesson 2: Serving a foundation model: memory, batching and cost33 min
- Lesson 3: Evaluating foundation models for Kestrel's own task34 min
- Lab: Message models evaluated, gated and costed186 min
- Incident: A science model that sounded right19 min
14Module 13: Specialisation5 parts · 257 min
- Module 13: Specialisation1 min
- Lesson 1: Nine specialist tracks, and a taste of each29 min
- Lesson 2: Depth in one track: tabular and risk24 min
- Lab: A paper result reproduced on model v2189 min
- Incident: A crop that always chose one face14 min
15Module 14: Responsible ML5 parts · 274 min
- Module 14: Responsible ML1 min
- Lesson 1: Fair and explained: who the model fails, and why29 min
- Lesson 2: Governance: records, privacy, attacks and saying no27 min
- Lab: Fairness, explanations and the model card for v2202 min
- Incident: False matches, no test, no monitoring15 min
16Module 15: Professional Practice5 parts · 274 min
- Module 15: Professional Practice1 min
- Lesson 1: Judgement: when to build, when to refuse28 min
- Lesson 2: Growth: read, reproduce, write up, publish24 min
- Lab: The Kestrel repository, ready for strangers206 min
- Incident: Right for the school, wrong for the student15 min
17Module 16: Capstone2 parts · 228 min
- Module 16: Capstone1 min
- Lab: One model that keeps working without you227 min
Training models, not just calling them, and then keeping them correct after they ship. The roadmap calls this "the infrastructure-and-training track, distinct from AI engineering": it "demands more mathematics and more systems depth", and it is for people who "want to build the systems that produce and serve models". Its central claim is the one this course is built around: "Training a model is a bounded problem with a clear finish. Keeping a model correct in production is unbounded, and that is where nearly all the engineering time actually goes."
You build one model: kestrel-late-risk at Kestrel, a UK parcel-delivery company with a Leicester hub, nine depots and about 210,000 parcels a day. The model predicts the chance that a parcel misses its promised day, so Kestrel can warn customers at 06:00 and send its few spare vans where they are needed. A prototype from last spring scored well, but nobody could rerun it, it trained on a column known only after delivery, and it has never seen a parcel from Doncaster, the depot that opened in August. You carry the real model from raw depot scans to a gated, benchmarked, monitored service, in one repository, with a decision record for every choice that needed one. Kestrel is fictional and its data is simulated; every lesson also opens with a real, cited case.
Who this course is for
You can write Python and you have seen a pandas DataFrame. The learner ranges from a beginner with some Python to a working engineer who has trained models but never run one in production.
Reviews
to review this course once you have finished it.