Status: next — starts the Monday after SWE 101’s finish line is met · Length: 6 weeks
How every week works: the four lanes, the rotation and the rules are in SWE 101 and don’t change. Notebook sections are numbered 102.1–102.6.
Then: SWE 103 — systems and interviews.
Why this course
SWE 101 left you with a deployed, tested flagship. This course gives it the thing that makes the target profile real: an AI feature you can prove works, with an evals harness and numbers you can quote. Almost nobody applying to junior AI roles has that. It’s the single highest-leverage gap left once graphs are under control.
Course finish line
- The flagship’s AI feature returns validated, structured output
- A golden set of at least 20 cases, a scorer, and an evals run in CI that fails when a prompt change makes results worse
- Numbers you can quote: pass rate, cost per request, p95 latency
- One write-up of the AI feature: what it does, how it’s evaluated, one failure you found through evals
- Linked lists, stacks, intervals and dynamic programming: about 30 problems logged, at least 20 cold
- Applications every Friday — 10 a week from here on — and every interview debriefed the same day
Week 1’s first job is the audit of this course’s lessons — K / H / D, as in SWE 101.
The six weeks
Week 1 — How models actually behave
- DSA — arrays, hashing, prefix sums: 004 group anagrams · 005 top K frequent elements · 007 product of array except self · 009 longest consecutive sequence. Refresh: prefix sum.
- Learn (core): how LLMs work · calling models
- Learn (optional): the AI engineer role · the model landscape · local and open models — measured on your own laptop
- Build: the AI feature’s skeleton — one model call through a port, streaming if the UI needs it, with a fake model adapter for tests.
- Apply: Friday hour, 10 a week.
- By Sunday: closed-book — why does a model hallucinate, mechanically?
Week 2 — Structure and prompts
- DSA — linked lists: 037 reorder list · 038 remove Nth node from end · 041 linked list cycle · 042 find the duplicate number · 043 LRU cache. Refresh: fast and slow pointers · linked list reversal.
- Learn (core): structured output · prompt engineering
- Learn (optional): RAG and embeddings — core instead if the flagship’s feature retrieves documents
- Build: the feature returns schema-validated output, and the prompt lives in version control as a file, not a string in the code.
- Apply: Friday hour.
- By Sunday: the AI feature works end to end in production.
Week 3 — Evals ⭐
- DSA — stacks: 021 valid parentheses · 024 generate parentheses · 025 daily temperatures · 026 car fleet · 027 largest rectangle in histogram. Refresh: monotonic stack.
- Learn (core): evals · AI engineering practice exercises
- Build: write the golden set — at least 20 real inputs with what a good answer looks like — before building any harness. Include the cases you expect it to fail.
- Apply: Friday hour.
- By Sunday: the golden set exists, and a scorer runs over it locally and prints a pass rate.
Week 4 — Evals in CI, and production behaviour
- DSA — intervals and greedy: 128 partition labels · 130 insert interval · 131 merge intervals · 132 non-overlapping intervals · 134 meeting rooms II. Refresh: overlapping intervals.
- Learn (core): reliability and plumbing · cost, caching and latency
- Learn (optional): resilience patterns
- Build: the evals run in CI and fail the build below a threshold. Add timeouts, a retry with backoff, and per-request cost and latency logging.
- Apply: Friday hour.
- By Sunday: the evals harness runs in CI, with numbers you can quote.
Week 5 — Tools, agents, and when not to use them
- DSA — dynamic programming I (one dimension): 101 house robber · 102 house robber II · 106 coin change · 108 word break · 109 longest increasing subsequence. Refresh: dynamic programming.
- Learn (core): tools and MCP · agents
- Learn (optional): multimodal
- Build: only if the feature genuinely needs a tool call — otherwise, use the week to raise the eval pass rate and record what changed it.
- Apply: Friday hour.
- By Sunday: closed-book — when is an agent the wrong architecture?
Week 6 — Safety, the write-up, and a timed mock
- DSA — dynamic programming II (two dimensions), then a mock: 111 unique paths · 112 longest common subsequence · 114 coin change II. Then one timed mock: two unseen mediums, 45 minutes each.
- Learn (core): safety and production — prompt injection and PII
- Build: the AI feature write-up, with the evals numbers and one failure the evals caught.
- Apply: Friday hour, with the write-up linked.
- By Sunday: every box in the finish line is ticked, or has a date. SWE 103 starts the following Monday.
Related
- SWE 101 — the series overview and the weekly rules
- The AI engineer track — the full set of lessons this course draws from
- Lesson quality
Software engineering
Infrastructure & security
Computer science
Hardware & engineering