How the learning sequence uses its sources
Review date: October 13, 2026. Yes, rohitg00/ai-engineering-from-scratch informs this plan. Its README, ROADMAP and Career Practice were reviewed, together with the selected lesson/project content linked below.
This is a curriculum/content review, not a whole-repository code audit. No upstream lessons are copied into these pages. The weekly assignments are our own small-project sequence with attribution links.
What was retained and adapted?
The reference emphasizes implementation, tests, failure inspection and one coherent project. We retain those practices and the retrieval lab's query-by-query comparisons. We narrow its broad curriculum into three starting-level routes, put selected resources beside assignments, include independent project/review time and separate optional depth.
The reference's phase timings are not used as completion guarantees. Its ROADMAP headline is roughly 332 hours; this plan uses its own 336/480-hour budgets and gates. Completing a source lesson or project does not establish professional experience.
Selected upstream content actually inspected
| Content | Selected use | Boundary |
|---|---|---|
| Git and Collaboration | B11: local versioning, branching and ignored files | Not verification of every setup lesson |
| Model evaluation | Supporting reference for splits, metrics and leakage | Primary scikit-learn/course readings lead the ML assignments |
| RAG | A15/A17 and C25–26: loading, chunking, embedding, storage, retrieval and generation | Broad cost/privacy/superiority claims are not adopted; task-specific comparison remains necessary |
| Retrieval Evaluation Lab | A16/A18: ranking validation, precision/recall, rank-sensitive metrics and per-query comparison | Project description/stage map inspected; runner and solution code not executed or audited |
| Tool schema design | A14/A19: descriptions and parameter constraints | Not independent verification of its benchmark claims |
| Agent loop | A19: registry, observations and stop/budget conditions | Selected concepts, not completion of its full prerequisite chain |
| Failure modes | A20/delivery: cascading errors, context loss and tool misuse | Paper/industry claims and all linked references not independently verified |
| Secrets and audit lesson | Background review only; not required weekly reading | Its incidents, universal patterns and policy thresholds are not adopted as verified standards |
| Career Practice | One project, prerequisite checks, reproduction and actual feedback | Its career-specific overlays are not required or an employment promise |
The links follow a moving main branch. Record the source revision/date you actually use for an exercise.
Primary education and documentation checked
| Source | Content accessed and relevance |
|---|---|
| CS50P | Course overview, weeks 0–9 pages and final-project requirements; direct selections in B |
| CS50 SQL | Overview and Querying/Relating/Designing pages; selected data foundations |
| ML specialization | Three-course description and supervised/unsupervised subject coverage; not protected assignments |
| fast.ai | Course overview and lessons 1–2 pages: coding prerequisites, model training/deployment and 2022 Part 1 recordings |
| HF LLM course | Introduction/prerequisites, Tokenizers, fine-tuning introduction and Datasets introduction |
| scikit-learn | Leakage/common pitfalls, cross-validation and probability calibration explanations/examples |
| SciPy bootstrap | Synthetic resampling example, variability and confidence intervals for C's statistical practice |
| PostgreSQL | Tutorial, transactions and window-function content |
| pytest, uv, FastAPI | Testing example, project/environment management, API first steps/request bodies/response models/errors/security and concurrency pages |
| Docker, Python CI | Application/container/Compose tutorial and build/test workflow guide |
| pgvector, MLflow | Vector search/persistence/update/delete and experiment/run/artifact tracking |
| MLOps Zoomcamp | README/access conditions and modules 2–6 descriptions: tracking, orchestration, deployment, monitoring and best practices |
| LangGraph, MCP | Persistence/interrupt content and protocol architecture; optional delivery concepts |
| Credential register | Official Microsoft AI-103/study-guide, retired AI-102, AWS ML Associate and Google ML pages rechecked; public scope/status only |
Verification limits
- Public pages and selected content were fetched; this is not proof that every nested link, video, code sample or course exercise works.
- Initial bare CS50
/python/0/and/sql/0/-style URLs returned 404. The provider's weeks index exposed the correct/weeks/…/paths; the selections in this plan were fetched at those corrected URLs. The index's simplified extraction omitted its lesson list, so its HTML was inspected. - Docker's old workshop URL redirected; the plan uses the fetched application tutorial. MCP's architecture URL redirected to a versioned documentation page. Pin protocol/SDK versions before implementation.
- No enrollment, checkout, cloud deployment or exam booking was performed. Paid/sign-in assignments, local price/appointments, assessment access and credential outcomes remain unverified.
- Source availability and subject coverage are not a “best-in-class” ranking or evidence of employment outcomes. Individual readiness needs actual artifacts and review.
Most linked resources are in English; the Spanish route explains the same assignments but does not promise translated provider material or Spanish exam availability.
Recheck broken links when encountered, version-sensitive APIs before implementation and official credentials before payment. Record a source/date change rather than silently assuming older exam codes remain current.