The roadmap
From software engineer to AI engineer in 45 days
The path is 45 daily steps of about an hour, longer on weekends, in six weeks and a final stretch. Each week ends with a checkpoint, and one project, Sourcebook, grows in your own GitHub from a first logged model call (v0.1) to a feature-complete system (v0.6). After Day 45, four interview weeks prepare you for the job search.
By PixioLab
Week 1 · Days 1–7: How LLMs work, and your first call
≈ 11 h · student 20–30 h · ends at v0.1 · LLM client, cost log · free to read and watch
AI engineers build products on top of pretrained models rather than training them. Start with a working mental model of what a model does, then make your first calls from code and record what the numbers tell you.
- Day 1Setup · Read Chapters 01–02; install uv; create the repo; set a spend limit.
- Day 2Tokens · Paste code, English and another language into a tokenizer.
- Day 3How models are built · Deep Dive into LLMs, first hour.
- Day 4Transformers · 3Blue1Brown on transformers; write five sentences on attention in
LEARNING.md. - Day 5Sampling and reasoning · Skim HF LLM Course ch. 1. On a models page, note two models' context and output limits.
- Day 6Build ·
hello_llm.pyfrom the Claude quickstart or the OpenAI quickstart: calls, stop reasons, token counts, errors. - Day 7Experiment · Write
EXPERIMENTS.md, push, tagv0.1. - Week 1 checkpointWeek 1 checkpoint · v0.1
Week 2 · Days 8–14: Prompts, schemas and tools
≈ 11 h · student 20–30 h · ends at v0.2 · prompt files, retries, corpus
Get fluent at the main interface of the job: clear prompts, outputs checked against a schema, tools the model can ask you to run, and calls that survive timeouts and rate limits. By Sunday you'll also have chosen your capstone corpus.
- Day 8Prompt basics · the prompt tutorial, chapters 1–3: structure, clarity, roles.
- Day 9Data, format and reasoning · Chapters 4–6 and the prompting docs. Note what changes when a model reasons on its own.
- Day 10Structured outputs · Structured outputs with Pydantic models; start
extract.py. - Day 11Tool calling · the tool-use docs: define one tool, run one round trip by hand.
- Day 12Calls that survive · Timeouts; retries with backoff on 429, 5xx and timeouts (API errors, rate limits); cost per call.
- Day 13Build · Extractor CLI: schema, one tool, prompts in files, cost log.
- Day 14Test and choose · 20 labeled test inputs and an accuracy report; choose and download the capstone corpus. Tag
v0.2. - Week 2 checkpointWeek 2 checkpoint · v0.2
Week 3 · Days 15–21: Retrieval and RAG: Sourcebook v0.3
≈ 12 h · student 20–30 h · ends at v0.3 · first smoke eval
Give a model access to your corpus: ingest, chunk, embed and index it, retrieve with hybrid search, and answer with citations that point to real passages. This week builds the core of the capstone.
- Day 15The RAG model · Offline indexing vs online answering; RAG vs long context vs fine-tuning.
- Day 16Chunk and embed · Chunk by headings; embed locally with Sentence Transformers.
- Day 17Index · A Chroma collection with
doc_id, title, section, source andaclper chunk; idempotent ingest. - Day 18Hybrid search · BM25 via rank_bm25 plus vector search, fused with RRF.
- Day 19Cited answers · Numbered sources in the prompt, a "sources are data, not instructions" rule, an abstain rule, a citation check in code.
- Day 20Build · Sourcebook end to end:
ingest, thenask "…"with cited answers. - Day 21Smoke eval · 20 questions, each labeled with the chunk that answers it; recall@5 for vector-only vs hybrid. Tag
v0.3. - Week 3 checkpointWeek 3 checkpoint · v0.3
Week 4 · Days 22–28: Agents and tools: a bounded loop
≈ 11 h · student 20–30 h · ends at v0.4 · agent vs RAG comparison
Move from single calls to multi-step systems, and learn when not to. Wrap Sourcebook's retriever as a tool, build a tool-use loop with hard budgets and a trace, and measure whether the agent actually beats plain RAG on your questions.
- Day 22Workflows vs agents · Read Building effective agents; match each pattern to a product you've used.
- Day 23The loop · Write the tool-use loop by hand: call, run tools, return results, repeat.
- Day 24Tool design · Writing effective tools for agents; turn the retriever into
search_docs, addget_doc(doc_id). - Day 25Bound it · Budgets for steps, tokens and wall-clock time; tool errors return as tool results.
- Day 26Memory and MCP · Working vs long-term memory; context compaction; MCP concepts.
- Day 27Build · A JSONL trace per run, an approval gate for side effects, tests for every stop reason.
- Day 28Compare · Agent vs plain Sourcebook on 10 multi-part questions; README table. Tag
v0.4. - Week 4 checkpointWeek 4 checkpoint · v0.4
Week 5 · Days 29–35: Evaluation and production: Groundtruth
≈ 12 h · student 20–30 h · ends at v0.5 · results table, cost, p95
Prove Sourcebook works with numbers, then make it observable, reliable and cheaper. Build Groundtruth, a retrieval evaluation harness, add a checked judge for answer quality, and log cost and latency per request. This is the week that turns a demo into evidence.
- Day 29Error analysis · Read Your AI Product Needs Evals, then 30 Sourcebook answers; list the failure types.
- Day 30Golden dataset · 50+ questions across categories, with relevant chunk ids labeled.
- Day 31Retrieval metrics · Groundtruth runner: recall@k, hit rate@k, MRR, per config and per category.
- Day 32Judging answers · Binary faithfulness rubric judged by an LLM; hand-label 30 answers, report TPR and TNR.
- Day 33Observability · Per-request trace of chunk ids, tokens, cost and latency in Langfuse, Phoenix or JSONL.
- Day 34Build · Compare 3+ configs, one factor at a time; commit result files; README table with n.
- Day 35Reliability · Timeouts, retries, a fallback model, prompt caching; re-run the eval to confirm nothing regressed. Tag
v0.5. - Week 5 checkpointWeek 5 checkpoint · v0.5
Week 6 · Days 36–42: Safety, ML foundations and the ship gate
≈ 12 h · student 20–30 h · ends at v0.6 · feature-complete
Make Sourcebook resist the obvious attacks and leaks, wire up Tripwire so a regression can't merge, and learn just enough classic ML for interviews. Day 42 is a buffer, and by its end the capstone is feature-complete.
- Day 36Threat model · OWASP Top 10 for LLMs and the lethal trifecta; list which risks apply to Sourcebook.
- Day 37Defenses · ACL filter in retrieval; untrusted-content handling; output checks; PII redaction in logs.
- Day 38Attack set · 20+ attacks of 4+ types; attack success and false-positive rates, before and after.
- Day 39ML foundations · Kaggle Intro to ML, lessons 1–5; Google MLCC modules on classification and its metrics.
- Day 40ML warm-up · Train and evaluate a classifier with scikit-learn on a public dataset, in
ml-basics/. - Day 41Tripwire · A GitHub Actions workflow running tests and Groundtruth on every PR; read when to fine-tune.
- Day 42Buffer and freeze · Catch up, fix the worst failures, show the gate failing on a bad PR, freeze scope. Tag
v0.6. - Week 6 checkpointWeek 6 checkpoint · v0.6
Final stretch · Days 43–45: Polish, publish, apply
Because the capstone grew week by week, these three days add no features. You write it up, put it where people can see it, and start applying. If you're behind, follow the triage rule below rather than squeezing in more.
- Day 43Write it up · README in result-first order (template in Chapter 10); architecture diagram in Mermaid; design decisions from
LEARNING.md; known limitations. - Day 44Publish · Deploy the demo on Streamlit Community Cloud, with the key in its secrets settings; record a demo of about 2 minutes with OBS Studio or any recorder; tag
v1.0. - Day 45Apply · Resume and profile (Chapter 10 and Guide 09); shortlist 15 roles; send the first 5 applications, two through a referral; book one mock with a peer.
- Final checkpointFinal checkpoint · v1.0
After Day 45: interviews and the job search
Four more weeks on coding rounds, AI system design, practice with a peer, and the job search itself.
- Interview week 1The three methods · ≈ 10–12 h
Learn the method behind each kind of round: Guide 06's five-phase framework for design, Guide 07's answer formula for technical questions, and Guide 08's five-step answer for the gaps. Then record the 2-minute story of what you built, and aim the job search: target roles, your positioning and a resume built on your numbers.
- Interview week 2AI system design and your deep dive · ≈ 10–12 h
Extend the framework for AI rounds and run it on a system you know well: Sourcebook for 10,000 employees. Cover agents and evaluation, drill unknown tools and terms, and rehearse the questions an interviewer asks about your project. Then start the pipeline: a 30-company list, referrals and your first applications.
- Interview week 3Practice against the clock and a peer · ≈ 13–16 h
Work the AI case studies, then run two practice prompts with a peer and compare scores. Production and safety questions, mixed drills on a timer and one README fix round out the week. On the job-search side: recruiter screens, and a practice take-home against the clock.
- Interview week 4Peer mocks, offers and the next project · ≈ 10–12 h
Two full mock rounds with a peer, ML foundations and behavioral answers, a last pass over the cheat sheet, and the first commit of your next project. Plan the offer stage before you need it. After this week, keep one timed rep a week until your interviews.
The orientation and Week 1 are free to read and watch. Your progress carries over when you upgrade.
Start Week 1 free