13-week schedule
Monday 12 October 2026 → Sunday 10 January 2027. About 5 hours a week: 4–4.5 hours of planned work plus a little slack for the primary-source reading.
Where you are
Loading…
A typical week
Three sessions of about 90 minutes beat one long Sunday, because spacing builds long-term memory.Spacing and retrieval practice are among the best-supported learning techniques: see Dunlosky et al., "Improving Students' Learning With Effective Learning Techniques" (2013).
- Session 1: lesson A (concept, prediction, lab, exercises).
- Session 2: lesson B.
- Session 3: lesson C, or the spaced review: re-do the exercises you missed earlier, read the lesson's primary source, and update
portfolio/LOG.md.
Behind schedule? Drop the stretch exercises in light lessons first. Keep the reviews and the checkpoints: they protect everything else you've learned.
Week by week
Milestones
| When | You can… |
|---|---|
| End of W3 | Explain everything that happens between sending a prompt and the last token, including the KV cache. |
| End of W5 | Do roofline napkin math, and have your first vLLM deployment running on a GPU (Checkpoint A). |
| End of W9 | Apply and measure quantization, speculative decoding, prefix caching and tensor parallelism (Checkpoint B). |
| End of W12 | Run a containerized, autoscaled, observable model service with a cost model (Checkpoint C). |
| End of W13 | Show a published capstone and pass a mock inference-engineering interview. |