The active track. The takeUforward roadmap broken into atomic notes, then every Codemia problem regrouped by shared architecture so there are 29 skeletons to learn instead of 217 problems to memorise. 50 of the 67 roadmap topics have notes.
The distributed-systems half of AI: retrieval architecture, agent topologies, inference cost, evaluation and safety. A twenty-part guide, of which six parts are fully written up as atomic notes so far.
The hands-on half — code I actually ran, then turned into notes. LangChain v1 and LangGraph end to end, plus the infrastructure around an agent: MCP servers, a model gateway, retrieval without a vector database, and evaluation.
create_agent, tools and the three-step execution loop, message types, structured output, middleware (summarization, human-in-the-loop) and a five-layer guardrail stack.
Open ↗
02
add_messages reducer, the ReAct back-edge, checkpointing with thread_id, interrupt() for human approval, multimodal RAG with CLIP, and sequential vs supervisor vs hierarchical multi-agent.
Open ↗
04
Preparing for Claude Certified Architect — Professional: 63 items, 120 minutes, 720 of 1000 to pass. I built a practice-exam engine that samples the real blueprint weights and teaches the topic on the spot when I get one wrong. Six of the seven domains have notes; Domain 6 is the open gap.
The machinery underneath: every note in one searchable file, the catalog, the running log, and the honest scorecard of what's still missing.