15. The Complete 24-Week Study Plan

This table reproduces the roadmap's week-by-week plan in full. Each week's "Learn" column maps to the corresponding chapter above for full depth.

Wk	Focus	Learn	Deliverable
1	Python fundamentals	Syntax, functions, collections, files, modules, exceptions.	CLI/data-processing mini app.
2	Production Python	Typing, Pydantic, async/await, httpx, logging, pytest.	Typed async service + tests.
3	FastAPI + PostgreSQL	REST, auth/RBAC, SQL, ORM, migrations, Docker.	Production API foundation.
4	AI/ML/LLM foundations	ML metrics, neural nets, transformers, tokens, context, embeddings.	Engineering explainer + small inference notebook.
5	Raw LLM APIs	Messages/instructions, model choice, streaming, errors, token accounting.	Chat/API endpoint without framework.
6	Structured outputs	JSON schema, Pydantic validation, extraction, reliability.	Document extractor v1.
7	Embeddings	Vectors, similarity, batching, semantic search.	Semantic search API.
8	pgvector + search quality	Vector indexes, metadata filters, lexical search baseline.	Search benchmark + relevance dataset.
9	RAG ingestion	Parsing, chunking, metadata, versions, jobs.	Async ingestion pipeline.
10	RAG retrieval	Dense retrieval, top-k, filters, query understanding.	RAG v1 with citations.
11	Advanced retrieval	BM25/keyword, hybrid search, reranking, query rewrite.	RAG v2 with hybrid + reranker.
12	RAG evaluation	Recall@k, answer correctness, faithfulness, citation checks.	RAG regression suite.
13	Tool calling	Schemas, typed tools, read operations, failures.	Business assistant with 3 read tools.
14	Safe write workflows	Approvals, idempotency, permissions, audit logs.	Add draft/create workflow with approval.
15	Agent fundamentals	State, routing, planning, retries, stopping.	Build small loop manually.
16	LangGraph	Graphs, persistence, interrupts, human-in-loop, streaming.	Stateful agent workflow.
17	Advanced workflows	Subgraphs, memory choices, durable execution, error recovery.	Long-running workflow with resume.
18	MCP	Client/server, tools/resources, auth/trust boundaries.	Build read-only MCP server + client.
19	Evals	Datasets, deterministic checks, LLM judge calibration, comparisons.	Automated eval harness.
20	Observability	Tracing, token/cost metrics, model/tool/retrieval telemetry.	Dashboard/traces for flagship app.
21	AI security	Prompt injection, RAG/tool threats, cross-tenant tests, OWASP.	Security test suite.
22	Production systems	Redis, queues, workers, object storage, rate limits, retries.	Async production architecture.
23	Cloud + CI/CD	Deploy, secrets, managed DB/storage, logs, eval/test gate.	Public staging deployment.
24	Flagship + interview	Polish README, architecture, demo, metrics, interview stories.	Production-grade portfolio release.

15.1 Weekly Rhythm (10–15 hours)

Session	Suggested Work
2 × 60–90 min	Concepts + official documentation.
3 × 90–120 min	Build feature, test, debug.
1 × 60 min	Write evals / security tests / refactor.
1 × 60 min	Document what you built and publish progress.

💡 If You Have More Time Increase project depth, not the number of frameworks. Add tests, metrics, deployment, failure handling and evaluation before moving to another library.

16. Portfolio Projects: Beginner to Flagship

Project 1 — Typed AI Document Extractor

FastAPI + Pydantic structured outputs.
Invoices/CVs/contracts/receipts.
Human correction UI.
Field-level eval set.
Persist documents and results.
Dockerized deployment.

Project 2 — Semantic Search Engine

500+ documents.
Embeddings + pgvector.
Lexical baseline.
Metadata/tenant filtering.
Recall@k evaluation.
Search latency metrics.

Project 3 — Production Knowledge RAG

Async ingestion pipeline.
Chunking experiments.
Hybrid retrieval + reranking.
Citations and source viewer.
Auth/RBAC + multi-tenant filtering.
RAG eval suite.
Observability and cost tracking.

Project 4 — AI Business / ERP Copilot

Read tools: customers, invoices, sales, inventory.
Analytical tool functions with safe backend queries.
Write actions only through approval workflow.
Agent/workflow state and audit logs.
Tool-selection/argument evals.
Prompt-injection and permission tests.

Flagship — AI Business Operations Platform

Next.js UI -> FastAPI API -> AI Orchestrator |- RAG / Retrieval |- Tool Gateway |- Stateful Workflows / Agent |- Evaluation Hooks -> PostgreSQL + pgvector -> Redis / Queue / Workers -> Object Storage -> Enterprise APIs / ERP -> Observability -> Cloud Deployment

Fig 16.1 — The flagship project is not a new stack; it is every prior project's components combined into one deployed system.

Flagship Feature Checklist

Auth + orgs + RBACStreaming responsesStructured outputsIngestion/statusRAG + citationsHybrid search + rerankingConversation stateTool callingHuman approval for writesAgent persistenceMCP (optional)EvalsTracingCost trackingPrompt/model versioningAudit logsPrompt-injection defensesRate limitingQueue/workersDockerCI/CDCloud deploymentArchitecture diagramLoad/security test reportREADME with measured results

📌 Sticky Note — Remember Ship one deep, measurable flagship product rather than several shallow chatbot demos. Every interviewer question in Chapter 18 maps back to something a real flagship project would force you to actually solve.

17. Production Architecture Reference

Fig 17.1 — Reference architecture for the flagship platform. Every box corresponds to a chapter already covered: FastAPI (Ch.3), AI Orchestrator/RAG (Ch.5–7), Tool Gateway (Ch.8, 10), Queue/Workers (Ch.13), Observability (Ch.11).

17.1 Cross-Cutting Controls

Authentication/authorization applied before retrieval/tool execution.
Tenant/resource IDs attached to every data access.
Request IDs/traces across API, model and tools.
Timeouts, budgets and rate limits.
Secrets never embedded in prompts.
Evals run in CI for important model/prompt/retrieval changes.
Audit logs for side effects.
Data-retention and privacy policy per provider/environment.

Chapters 15–17 Summary

The 24-week plan sequences every phase chapter into weekly deliverables, moving from ordinary backend work to a full production AI platform.
Four escalating portfolio projects (extractor → search → RAG → ERP copilot) build directly toward the flagship platform.
The flagship project is not new material — it is the integration of every prior chapter's components into one deployed, observable, secured system.
The reference architecture diagram is a map back onto the chapters: every box was already taught somewhere in Chapters 3–14.