Phases & exit gates¶

A phase ends when its gate is met, not when the calendar says so. If a gate isn't met, use the next light week or buffer weeks 25–26. Never drop a P0 item.

Phase 0 — Baseline¶

Week 0 · Calibrate starting level and get the toolchain working so week 1 starts with building, not setup.

Gate (exit criteria):

Baseline SD mock scored with the rubric
3 DSA mediums attempted with times logged
AI self-assessment scores + skip list written to docs/log/baseline.md
uv, Docker, Ollama and API keys verified with a smoke-test call

Phase 1 — Foundations¶

Weeks 1–4 · Refresh modern Python/Java/LLM foundations and start the capstone with structured outputs, tools and workflow patterns.

Gate (exit criteria):

Capstone repo with FastAPI /triage returning validated structured output and a typed tool loop
Orchestrator-workers + evaluator-optimizer feature working with 10 golden examples
ADRs 001-004 written; AGENTS.md in capstone repo
Arrays/hashing, two pointers, sliding window, stack: 20+ problems solved
interview-1 mock scored >= 2.5/4 average on the rubric

Phase 2 — RAG + Evals¶

Weeks 5–8 · Ship hybrid RAG with reranking and an eval-driven workflow: error analysis, aligned judges and CI gates with tracing.

Gate (exit criteria):

Hybrid retrieval beats dense-only on the golden set (hit rate@5 reported)
Failure taxonomy from >= 100 traces; LLM judge TPR/TNR >= 0.8 against hand labels
Eval job in CI failing on threshold regressions; Langfuse traces for every request
Design doc v1 reviewed; ADRs 005-008
interview-2 mock scored >= 2.75/4

Phase 3 — Agents + Protocols¶

Weeks 9–12 · Build a durable LangGraph orchestrator with Pydantic AI sub-agents, MCP servers in Python and Spring AI, HITL and memory.

Gate (exit criteria):

LangGraph orchestrator survives kill/restart mid-run and resumes from checkpoint
HITL approval gates every side-effecting tool call
Python MCP server + Spring AI MCP server both called by the orchestrator over Streamable HTTP
Written multi-agent platform design + influence plan artifact
interview-3 mock scored >= ¾

Phase 4 — Production & AI Architecture¶

Weeks 13–16 · Make the capstone production-grade (guardrails, routing, cost, observability, SLOs) and produce Staff-level architecture/strategy artifacts.

Gate (exit criteria):

promptfoo OWASP agentic red-team suite in CI with no high-severity failures
Routing cuts cost per resolved incident >= 30% without eval regression
SLOs + burn-rate alerts + online evals live; C4 + ADRs 013-016
Strategy doc, tech-debt proposal and AI adoption plan written
interview-4 mock scored >= ¾

Phase 5 — Scale & Depth¶

Weeks 17–20 · Go deep: DSPy/GEPA, inference serving, fine-tuning trade-offs, advanced RAG, distributed consensus, free-threaded Python, JVM concurrency.

Gate (exit criteria):

DSPy vs hand-prompt vs fine-tune ablation written with numbers
Self-hosted model serving experiment with cost/quality comparison
Gossip Glomers broadcast challenge passing; Raft explained in < 5 min
Mentoring plan + talk/blog draft done
interview-5 mock scored >= 3.25/4

Phase 6 — Capstone + Interview Loop¶

Weeks 21–24 · Ship capstone v1.0 on Azure with write-ups, and run full interview loops until scores are consistently strong.

Gate (exit criteria):

Capstone deployed on Azure Container Apps + Foundry with passing smoke evals
2 blog posts, README, demo video, arc42 doc and ADR set 001-024 published
STAR bank with 12+ stories
interview-6 mock scored >= 3.5/4 on every dimension

Phase 7 — Buffer¶

Weeks 25–26 · Catch up on slipped work and pick a specialization for the next 6 months.

Gate (exit criteria):

No P0 task left incomplete in progress.yml
Specialization spike written up
Final retro + next-6-month plan published