Extend the same project into operational practice

For: learners who already demonstrate software, ML evaluation and retrieval foundations. Additional proposed budget: 24 weeks × 15 hours = 360 hours: 16 shared weeks (240 hours) plus one eight-week focus (120 hours). Allocate about 4 hours reading, 9 implementation/debugging and 2 review/communication weekly.

This is optional depth, not a prerequisite for A/B/C completion. Select it after capability review, or choose individual exercises with an explicit additional budget. It does not award professional experience.

Before you start

Independently show a tested API, SQL joins/window functions, a containerized release, a clean ML split, a held-out retrieval experiment and one diagnosed CI/runtime failure. Repair missing foundations in A/C first.

Keep the same public/synthetic learning-support workbench. Add a mock external question API and human-reviewed draft—not real integrations or writes. Use the existing stack; LangGraph and MCP illustrate state and protocol concepts, not a requirement to collect frameworks. Pin compatible versions and record them.

Shared practice, one assignment per week

Week	Selected reading	Build and show
1	uv projects and pytest	A typed package, locked environment, tests and clean-clone commands.
2	FastAPI concurrency	Compare blocking and nonblocking I/O; test timeout and cancellation without claiming CPU parallelism.
3	FastAPI request validation and security overview; PostgreSQL transactions	Add application-level access checks, synthetic identity-isolation tests and a failed migration/rollback exercise.
4	MLOps orchestration: repeatable parameterized work	Make repeated requests idempotent; cap worker/queue/retry use and test throttling.
5	pgvector: search, filters and index tradeoffs	Rebuild a persistent index; compare lexical/vector search and test update/delete/access filtering.
6	MLflow tracking	Version prompts/models/data and run baseline/candidate evaluation. Calibrate any model judge against human labels.
7	LangGraph persistence: checkpointers vs stores	Resume after process restart with durable storage. Test session isolation, retention and deletion; in-memory examples alone do not show durability.
8	MCP architecture: host/client/server and data layer	Connect one local read-only tool. Pin protocol/SDK versions and test malformed and unauthorized calls.
9	LangGraph interrupts	Pause a draft for external review; bind application approval to identity and payload. Deny changed/replayed approval. An interrupt is not authorization.
10	Agent failure modes	Trace a failed tool and budget exhaustion. Redact logs; explain latency and token/cost counting without double counting.
11	MLOps monitoring	Predeclare load/duration and cost/latency/error budgets; run a load experiment and simulate an outage.
12	MLOps best practices: versioned infrastructure and delivery	Recreate one sandbox from reviewed configuration; if using IaC/cloud, review the plan and verify targeted teardown.
13	Docker application tutorial: multi-container application	Recreate API/data/worker processes locally; test restart behavior and readiness. No cluster is required.
14	Python CI and MLOps best practices	Make CI block a critical seeded regression; rehearse rollback and record an incident timeline.
15	Upstream Career Practice: feedback and ownership	Integrate the mock API; recover a broken contract, explain alternatives and improve one issue from actual review feedback.
16	Shared review rubric and capability guide	Have someone reproduce one failure/result. Present a design walkthrough and short recovery brief; mark review unavailable if needed.

These are focused readings/labs, not completion of every course or documentation site. Some source examples use different frameworks/clouds; transfer the concept to your project, do not adopt all their infrastructure.

Choose one eight-week focus

Each row covers two weeks: read/experiment first, implement/review second. Do not combine all focuses within the same 120 hours.

Stateful workflows

Weeks	Selected resource	Assignment
17–18	LangGraph persistence	Separate conversation checkpoints from long-lived data; test restart and isolation.
19–20	MCP architecture	Add one tool contract; document compatibility and denied/malformed requests.
21–22	LangGraph interrupts, failure modes	Test lost state, stale approval, timeout and deletion. Keep permissions outside model output.
23–24	Project evidence rubric	Compare direct vs framework flow on the same tasks; show quality, latency, cost and recovery differences.

Data and model operations

Weeks	Selected resource	Assignment
17–18	MLOps orchestration	Backfill a missing period; prove duplicate handling and dataset lineage.
19–20	Experiment tracking	Compare two versioned models with unchanged splits; explain promotion criteria.
21–22	Deployment and monitoring	Add batch scoring, test feature consistency and simulate drift vs corrupted data.
23–24	MLOps best practices	Reproduce training-to-serving and rollback; show what the operational test does not establish.

Evaluation and design communication

Weeks	Selected resource	Assignment
17–18	Retrieval Evaluation Lab	Diagnose per-query regressions hidden by an aggregate improvement.
19–20	Cross-validation, answer rubric	Write a controlled offline experiment with untouched data, slice results and stop criteria.
21–22	Upstream Career Practice	Run a peer design review and simulated incident; record actual feedback and an implemented improvement.
23–24	Evidence record	Present the same decision to technical and nontechnical audiences; distinguish measured results from assumptions.

Exit assessment

Use demonstrated / needs practice / not assessed / outside current scope for API/data contracts, state/tools, evaluation, load budgets, recovery and communication. Attach run commands and actual test results. A reviewer should recreate the sandbox and one failed/recovered operation.

An offline experiment is not live A/B testing; a local container is not production operations; a practice discussion does not certify a language level. Keep customer information private and obtain authorization before sharing professional artifacts or operating outside the synthetic lab.

Next: focused resources and verification limits.