Notes from the field on Data Science and Generative AI — production RAG, multi-agent systems, LLMs, and MLOps. Written from real shipped work, not hype.
How LangGraph runs graphs in parallel — the Pregel superstep model, fan-out/fan-in with reducers, the Send API for map-reduce, async nodes, and running whole subgraphs concurrently. With verified code and diagrams.
Break one bloated chat invoke into a session init that loads all static context once and a cheap per-turn process step — cutting latency and complexity with prompt caching in LangChain + Azure OpenAI.
Why single-agent RAG breaks down at scale, the multi-agent architecture that fixed it on LangGraph + Azure OpenAI, and the changes that lifted retrieval relevance 25% and cut latency 45%.