Three plans depending on how much runway you have. All of them assume ~2 hours/day on weekdays and a longer weekend block. Whichever plan you pick, the method is the same:
- Read the crash course (
README.md) in each topic first - it's the compressed theory. - Self-quiz with
questions.md- read the question, answer out loud before opening the collapsible answer. Speaking your answers is the single highest-leverage habit in interview prep. - Type out the coding challenges yourself - don't read the solutions first. The interview is a blank editor, not a multiple-choice test.
- Practice system design on a whiteboard or doc, talking through the framework before checking the case study.
- The night before any interview: CHEATSHEET.md.
Triage plan. Skip depth, maximise coverage of what's most likely to be asked.
| Day | Focus | Material |
|---|---|---|
| 1 | LLM fundamentals | 02-llm-fundamentals crash course + Basic/Intermediate questions |
| 2 | RAG + prompting | 04-rag-and-retrieval + 03-prompt-engineering-and-context crash courses, skim questions |
| 3 | Agents + evals | 06-agents-and-tool-use + 07-evaluation-and-observability crash courses + Basic questions |
| 4 | Coding reps | 12-coding-challenges: 01 attention, 03 sampling, 08 mini-RAG - implement before peeking |
| 5 | System design | 11-ai-system-design framework + the case study closest to the company's product |
| 6 | Production + safety | 08-inference-and-production Basic/Intermediate + 09-safety-security-and-responsible-ai crash course |
| 7 | Simulate + rest | 13-interview-process-and-behavioral - prep 5 STAR stories; evening: CHEATSHEET.md only |
Skip if you must: 10-multimodal (unless the role touches vision/audio), Advanced questions everywhere.
One theme per week; coding challenges spread throughout so implementation skills compound.
- Days 1-2: 01-ml-and-dl-foundations - full pass.
- Days 3-5: 02-llm-fundamentals - full pass, including Advanced.
- Weekend: challenges 01 attention, 02 BPE, 03 sampling, 04 positional encodings, 05 layernorm/softmax.
- Days 1-2: 03-prompt-engineering-and-context - full pass.
- Days 3-4: 04-rag-and-retrieval - full pass.
- Day 5: 05-fine-tuning-and-alignment - crash course + Basic/Intermediate.
- Weekend: challenges 08 semantic search/RAG, 09 chunking; finish fine-tuning Advanced questions.
- Days 1-2: 06-agents-and-tool-use - full pass.
- Day 3: 07-evaluation-and-observability - full pass. Do not skip this; it's the most senior-signalling topic in the repo.
- Days 4-5: 08-inference-and-production - full pass.
- Weekend: challenges 06 KV cache, 10 agent loop, 11 rate limiter, 12 eval metrics.
- Day 1: 09-safety-security-and-responsible-ai + 10-multimodal crash courses.
- Days 2-3: 11-ai-system-design - framework, then 3 case studies as mock interviews: 45 minutes talking into a doc before reading the solution.
- Day 4: 13-interview-process-and-behavioral - write your 5-7 STAR stories down.
- Day 5: challenges 07 mini-GPT forward, 13 streaming parser (the hard ones).
- Weekend: full mock loop - one coding challenge cold, one design prompt from the rapid-fire list, behavioural answers out loud. Hand the kit to a friend, or record yourself and score it a day later. Then CHEATSHEET.md.
Weeks 1-4: same as the 4-week plan, at half pace - and build while you learn:
- After Week 2's material → build a small RAG app over your own notes/docs with an eval harness (even 30 golden questions). This single project teaches more than any tutorial.
- After Week 3's material → add an agent with 2-3 tools to it, plus tracing.
Weeks 5-8:
| Week | Focus |
|---|---|
| 5 | Depth: re-do every Advanced section across topics 02, 04, 05, 06, 08. Read 5-6 foundational papers from resources (Attention, InstructGPT, LoRA, DPO, ReAct at minimum). |
| 6 | Projects: polish one portfolio project to "shows evals + error analysis + tradeoff writeup" standard (see project ideas in 13-interview-process-and-behavioral). |
| 7 | System design: all 8 case studies in 11-ai-system-design as timed mocks. All 13 coding challenges done cold. |
| 8 | Interview simulation: mock loops with a friend, or alone with the self-mock protocol; behavioural stories rehearsed; company-specific research; CHEATSHEET.md passes. |
- Spaced repetition beats rereading. Second pass on a topic 3-4 days after the first, third pass a week later. The
questions.mdfiles are already flashcard-shaped - question first, answer hidden. - Track your misses. Keep a running list of questions you fumbled; re-quiz only those on later passes.
- Explain to a human (or a rubber duck). If you can't explain the KV cache to a non-ML friend, you don't own it yet.
- Do the numbers by hand once. GPU memory maths, KV cache size, cost-per-request token maths - each done once on paper sticks forever.