The roadmap¶
Learn by doing. Vendor-neutral where it matters — dual-run local Ollama (free, private) against a frontier API (Claude/GPT/Gemini) so you feel the trade-off, not just read about it. 26 weeks, ~17–18 h/week. Verified 26 Sep 2026 — if a link looks stale, trust the tool's own docs over this page.
Legend: P0 blocks everything after it · P1 core, do it early · P2 depth, do it when a project needs it · ∥ can run in the background (a talk, a chapter) while you build something else · 🚢 a real ship point — working software a peer could use, not a checkbox.
Spine¶
Five resources that pay off across the whole plan, not one topic. Read them once early, then dip back in.
| Resource | Use it for |
|---|---|
| Anthropic — Building effective agents | The patterns ladder every agentic-ai topic assumes you know |
| Chip Huyen — AI Engineering | The systems view: evals, RAG, agents, deployment, in one head |
| Kleppmann & Riccomini — DDIA, 2nd ed. | The distributed-systems ground truth under every design decision |
| Percival & Gregory — Architecture Patterns with Python (Cosmic Python, free) | The Python architecture vocabulary (ports, adapters, UoW) used across the capstone |
| Will Larson — StaffEng | What "Staff" actually means before you try to act like one |
Phase 0 — Setup (this weekend, ~3–4 h)¶
# Toolchain
curl -LsSf https://astral.sh/uv/install.sh | sh
uv python install 3.14 3.14t
brew install --cask docker ollama # or your platform's equivalent
# Local models — pick one small + one embedding model
ollama pull qwen3:8b # or llama3.1:8b
ollama pull nomic-embed-text
# Repo
git clone <your-capstone-repo> agentic-ops-copilot && cd agentic-ops-copilot
uv init && uv add pydantic pydantic-ai fastapi langgraph
docker compose up -d postgres # capstone data layer, from week 1
Then the week-0 baseline: a system-design mock, 3 DSA mediums, an AI self-assessment, a recorded 2-min talk. It produces the skip list that makes week 1 start on building, not re-learning.
The tracks¶
Nine chapters, run mostly in parallel, not in series. A–D are sequential (each needs the last); E–I run every week alongside them at the hours shown. Full day-by-day schedule: weekly rhythm → all 26 weeks → phase gates.
A — Foundations (wk 1–4, P0)
Do: modern Python/Java tooling, LLM fundamentals, prompting, tool calling, agent patterns. ∥ read the Spine while you set up.
Deliverable: a typed tool-calling agent with structured output, in a repo with AGENTS.md.
→ Agentic AI ch. 1–5 · Python · Java & Spring AI
B — RAG as a spectrum (wk 5–8, P0) Do: chunking → hybrid search + rerank → pgvector/Qdrant → agentic/GraphRAG only once plain RAG is measured and beaten. Deliverable: hybrid RAG over your own corpus, with an ablation report (chunking × hybrid weights × reranker) and numbers, not vibes. → Agentic AI ch. 6–9 · P2: hybrid RAG with ablation
C — Evals & observability (wk 5–8, P0, parallel with B) Do: error analysis on real traces before you write a single eval. Only then promptfoo/DeepEval/Ragas, then Langfuse + OTel GenAI tracing. Deliverable: a failure taxonomy from ≥100 traces, an eval suite gating CI, and traces on every request. 🚢 SHIP POINT 1 (end of week 8): hybrid RAG + eval harness + tracing, running end to end. Show it to someone. → Agentic AI ch. 10–12
D — Orchestration & protocols (wk 9–12, P0)
Do: LangGraph for the durable, branching, checkpointed orchestrator; Pydantic AI for typed sub-agents; MCP for tools (one server in Python, one in Spring AI); durable execution + HITL before autonomy.
Menu (choose on merit): LangGraph vs Pydantic AI vs vendor SDKs — the framework choice guide has the real trade-offs, not a default answer.
Deliverable: the orchestrator survives kill -9 mid-run and resumes from checkpoint.
→ Agentic AI ch. 13–21 · P3: MCP agent + failure analysis
E — Production hardening (wk 13–16, P0) Do: OWASP LLM/Agentic Top 10 + red-team in CI, LiteLLM routing with budgets, cost/latency work, SLOs, Foundry/AgentCore deploy. 🚢 SHIP POINT 2 (end of week 16): capstone is guardrailed, routed, observable and deployed — production-grade, not a demo. → Agentic AI ch. 22–24, 27, 29 · P4: LLM gateway & guardrails
F — System design (wk 1–20, P0, ~2.25 h/wk, parallel to everything) Do: one topic + one case study per week, following the recommended path; estimation and scalability come first, consensus last. → System Design · AI System Design
G — Architecture & DSA (wk 1–24, P0/P1, ~3.5 h/wk combined, parallel) Do: one ADR/week on the capstone's real decisions; DSA patterns with re-solves on days ⅓/7/21 from the problem tracker. → Architecture · DSA
H — Staff+ & Communication (wk 1–26, P1, ~4 h/wk combined, parallel) Do: one Staff artifact every ~2 weeks (strategy doc, tech-debt proposal, mentoring plan…); 30 min/day of graded English/communication drills. → Staff+ Skills · Communication & English
I — Depth & specialisation (wk 17–20, P1) Do: DSPy/GEPA optimisation, self-hosted serving (vLLM/SGLang), a real fine-tuning ablation, Gossip Glomers for distributed-systems muscle. Deliverable: a go/no-go memo on fine-tuning, backed by numbers. → Agentic AI ch. 16, 25–26 · P5: optimise & serve · P6: distributed systems lab
🚢 SHIP POINT 3 — Capstone v1.0 (end of week 24): deployed on Azure, write-ups published, full interview loops run until scores hold steady. Weeks 25–26 are buffer: catch-up, then pick what's next.
Exit criteria (gate, not vibe)¶
Every phase has a measurable gate — thresholds, not "I read the chapter." → Phases & exit gates
All weeks¶
Same plan, mechanical: one row per week with its build, checkpoint and theme. → Weekly rhythm · Today
| Week | Starts | Phase | Theme | Build | Checkpoint |
|---|---|---|---|---|---|
| W00 | 21 Sep | 0 · Baseline | Baseline & environment (light) | Dev environment ready (uv, Docker, Ollama, API keys) + baseline scores recorded in docs/log/ | |
| W01 | 28 Sep | 1 · Foundations | Agent patterns, structured outputs, SD framework | Capstone repo skeleton: uv + FastAPI + Pydantic structured-output /triage endpoint | |
| W02 | 05 Oct | 1 · Foundations | Tool calling, API design, two pointers | Capstone: typed tool calling loop (runbook search, service status, ticket create) | |
| W03 | 12 Oct | 1 · Foundations | Context engineering, workflows, sliding window | Capstone: routing + prompt-chaining workflow with history compaction | |
| W04 | 19 Oct | 1 · Foundations | Orchestrator-workers, AI-assisted dev, stack — interview-1 | Capstone: orchestrator-workers 'draft postmortem' with evaluator-optimizer loop + 10 golden examples | interview-1 |
| W05 | 26 Oct | 2 · RAG + Evals | RAG fundamentals, caching, hexagonal | Capstone: RAG v0 — ingest runbooks/postmortems into pgvector, naive dense retrieval behind a Retriever port | |
| W06 | 02 Nov | 2 · RAG + Evals | Hybrid search + reranking (light week) (light) | Capstone: hybrid retrieval (BM25 + dense + metadata filters) with cross-encoder reranking | |
| W07 | 09 Nov | 2 · RAG + Evals | Evals I — error analysis, LLM-as-judge | Capstone: eval harness — error analysis on 100 traces, failure taxonomy, LLM-as-judge aligned to labels | |
| W08 | 16 Nov | 2 · RAG + Evals | Evals II in CI + observability — interview-2 | Capstone: promptfoo/DeepEval/Ragas in CI with thresholds + Langfuse tracing | interview-2 |
| W09 | 23 Nov | 3 · Agents + Protocols | LangGraph orchestrator | Capstone: LangGraph orchestrator (state, nodes, conditional edges, Postgres checkpointer) | |
| W10 | 30 Nov | 3 · Agents + Protocols | MCP + Pydantic AI sub-agents | Capstone: Python MCP server (runbooks, service status) + Pydantic AI sub-agents inside LangGraph nodes | |
| W11 | 07 Dec | 3 · Agents + Protocols | Durable execution, HITL, multi-agent | Capstone: human-in-the-loop approval (interrupt/resume) + Spring AI MCP server wired in | |
| W12 | 14 Dec | 3 · Agents + Protocols | Protocols, memory, vendor SDKs (light) — interview-3 (light) | Capstone: A2A/AG-UI spike + Mem0 memory prototype + vendor SDK comparison notes | interview-3 |
| W13 | 21 Dec | 4 · Production & AI Architecture | Guardrails & security | Capstone: guardrails layer (input/output/tool), prompt-injection red-team suite in CI | |
| W14 | 28 Dec | 4 · Production & AI Architecture | Cost, latency, routing | Capstone: LiteLLM routing (cheap → strong escalation), semantic + prompt caching, streaming latency budget | |
| W15 | 04 Jan | 4 · Production & AI Architecture | Managed platforms & production checklist | Capstone: Microsoft Foundry Agent Service spike + production-readiness checklist pass | |
| W16 | 11 Jan | 4 · Production & AI Architecture | Architecture docs & AI-era leadership — interview-4 | Capstone: C4 + arc42 doc set, observability dashboards, SLOs live | interview-4 |
| W17 | 18 Jan | 5 · Scale & Depth | DSPy + Raft + free-threaded Python | Capstone: DSPy/GEPA-optimised triage classifier vs hand prompt, measured with the eval harness | |
| W18 | 25 Jan | 5 · Scale & Depth | Inference serving (light) (light) | Capstone: self-hosted vLLM (or Ollama) serving experiment with quantised model behind LiteLLM | |
| W19 | 01 Feb | 5 · Scale & Depth | Fine-tuning trade-offs, GraphRAG | Capstone: LoRA/QLoRA fine-tune experiment (Unsloth/TRL) vs prompt+DSPy baseline; GraphRAG spike | |
| W20 | 08 Feb | 5 · Scale & Depth | Consolidation — interview-5 | Capstone: advanced RAG decision implemented (or rejected with evidence); talk/blog draft | interview-5 |
| W21 | 15 Feb | 6 · Capstone + Interview Loop | Capstone polish + STAR bank I | Capstone: deploy to Azure Container Apps + Foundry; end-to-end demo script | |
| W22 | 22 Feb | 6 · Capstone + Interview Loop | Write-ups + STAR bank II | Capstone: README, demo video, blog post #1 (eval-driven agent development) | |
| W23 | 01 Mar | 6 · Capstone + Interview Loop | Full mock loops | Capstone: blog post #2 (guardrails/architecture), final ADR review, portfolio page | |
| W24 | 08 Mar | 6 · Capstone + Interview Loop | Final loop — interview-6 | Capstone v1.0 shipped: repo, deployed demo, 2 blog posts, ADR set, arc42 doc | interview-6 |
| W25 | 15 Mar | 7 · Buffer | Buffer I — catch-up + specialization choice (light) | Catch up on any unfinished capstone increments; choose a specialization (serving, evals platform, or multi-agent) and scope a 2-week spike | |
| W26 | 22 Mar | 7 · Buffer | Buffer II — catch-up + next plan (light) | Finish specialization spike; publish final retro and next-6-month plan |