C: deepen the data, training and operational foundations

Start here if: you already build tested APIs, use Git/CI, query SQL and deploy services. Budget: 32 weeks × 15 hours = 480 hours: roughly 6 hours learning, 7 implementation/evaluation and 2 reviewing each week.

This journey deliberately goes deeper into machine learning and data than A. Training, data quality and repeatable pipelines are central—not an optional notebook beside a language-model feature.

Entry gate

Show a service you can explain, a SQL join/window query, a reproducible container and a CI failure you diagnosed. Fill software gaps with A first. Software experience does not establish ML understanding.

Use the same learning-support workbench: version question/topic events, train a classifier and compare documentation-answer approaches. These are three outcomes of one system, not three unrelated projects.

Phase 1 — ML foundations and statistical judgment

Weeks 1–6 / 90 hours. Complete the ML specialization's required work if seeking its certificate; extend the phase if it exceeds this allocation. Supplement a mathematical gap rather than assuming coding familiarity replaces it.

Weeks	Selected resource and focus	Assignment and evidence
1–2	ML specialization: regression, classification, loss and optimization	Explain features, vector/matrix shapes and gradients in a small model. Compare heuristic and linear baselines.
3–4	ML specialization: neural networks, trees and model selection; cross-validation	Choose random, grouped or temporal splits appropriately. Tune on development data only; compare model errors by slice.
5–6	ML specialization: unsupervised/recommendation material and remaining assignments; common pitfalls; SciPy bootstrap example: resampling and intervals	Explain leakage, imbalance, missing values and thresholds. Estimate variability by repeatedly sampling synthetic independent observations; document the interval and assumptions. Gate C1: baseline report, split manifest and justified limitations.

Phase 2 — Turn data into a recoverable product

Weeks 7–12 / 90 hours. Start with a local batch pipeline. Scheduling tools are optional until the command-line pipeline works.

Weeks	Selected resource and focus	Assignment and evidence
7–8	CS50 SQL: Designing, PostgreSQL transactions and window functions	Define keys, constraints and event history. Test failed transactions and calculate rolling topic counts.
9–10	MLOps orchestration module: parameterized scripts, dependencies and backfills	Build incremental ingestion; rerunning the same batch must not duplicate records. Quarantine malformed records and recover a failed run.
11–12	HF Datasets introduction: loading, slicing and processing	Version dataset snapshots, transformations and schema changes. Gate C2: backfill, deletion, quality checks and lineage reproduce from source to training set.

Phase 3 — Train and adapt a small model

Weeks 13–18 / 90 hours. The fast.ai page identifies Part 1 recordings as 2022; use current library documentation and compatible pinned versions. Selected lessons are not full-course completion.

Weeks	Selected resource and focus	Assignment and evidence
13–14	fast.ai lesson 1: training/using a model; lesson 2: your dataset and deployment	Train a small model, save weights and compare training/validation behavior. Explain overfitting and the compute budget.
15–16	HF Tokenizers and fine-tuning chapter introduction, then its preprocessing/training sections	Run a small text-classification training exercise on versioned data. Record tokenizer, model, seed, metrics and resource use.
17–18	MLflow tracking: runs, experiments and artifacts; probability calibration: calibration curves; revisit leakage pitfalls	Compare heuristic, linear/tree and adapted-model results on the same untouched split. Check whether predicted confidence matches observed frequencies; fit any calibration on development data only. Gate C3: reproducible training and honest baseline comparison.

Use CPU-sized or small models first. Fine-tuning the application is not mandatory if it is unjustified; the controlled training exercise still provides learning evidence. GPU or API spend needs a cap and cleanup plan.

Phase 4 — Connect training to serving

Weeks 19–24 / 90 hours. Select the indicated MLOps Zoomcamp topics—not the complete additional nine-week course. Translate its examples to your existing project rather than adopting every demonstrated service.

Weeks	Selected resource and focus	Assignment and evidence
19–20	Experiment tracking module: artifacts, lineage and model versions	Track training runs and register/version a candidate. Test feature consistency between training and serving.
21–22	Deployment module: batch and online serving; best practices: tests and CI	Serve one model by API and batch job. Block a seeded regression in CI; rehearse a failed release and rollback.
23–24	Monitoring module: health, quality and drift	Simulate missing/shifted data; distinguish drift from a pipeline bug. Gate C4: training-to-serving pipeline, alerts/report and recovery evidence.

Phase 5 — Compare retrieval and model adaptation

Weeks 25–28 / 60 hours.

Weeks	Selected resource and focus	Assignment and evidence
25–26	Upstream RAG lesson and pgvector: search and persistence	Build lexical/embedding retrieval; enforce document access filters outside the model. Test updates, deletion and denied access with synthetic identities.
27–28	Retrieval Evaluation Lab: query comparisons; shared answer rubric	Gate C5: compare prompt-only and RAG on one frozen answer task. Compare adaptation only if trained for that same task; classifier scores are not answer-quality scores. Report slice errors, cost and complexity.

The upstream lab has POSIX-dependent runner features; use WSL where needed or implement the metric exercises independently. Record which checks actually ran.

Phase 6 — Consolidate and explain

Weeks 29–32 / 60 hours.

Weeks	Selected resource and focus	Assignment and evidence
29–30	Capability guide, MLflow tracking and optional credentials	Repair the weakest gate. Optionally map one relevant exam syllabus to gaps; exam preparation is not automatically included in 60 hours.
31–32	Upstream Career Practice: feedback and reproduction; review practice	Gate C6: a reviewer recreates one data failure, one training result and one serving recovery. Write three connected case studies with actual results and limitations.

Optional depth, kept separate

Distributed processing, streaming, feature stores, Kubernetes, recommendation, forecasting and vision are electives. Add one when a clear dataset/workload or learning question justifies it; record additional time and cost. Do not rebuild infrastructure to hide weak data or evaluation.

For stateful workflows or more operational practice, the delivery extension adds 360 proposed hours if taken in full. It is not hidden inside C's 480 hours. Certificates and project uptime do not award professional experience.