A post-sales chatbot built with LangChain and LangGraph where a decision model picks the subagent per message: architecture, code, fallback, A/B and failures.
I had Microsoft-Decision-1 pick the subagent and LLM in a post-sales chatbot: harness, calibration, cost guard, and what 1,000 synthetic conversations showed.
How to structure a React SPA in vertical slices by domain: public API per slice, data from route loaders, and an architecture test that blocks shortcuts in CI.
How an LLM base model becomes ChatGPT: supervised fine-tuning, tool use, hallucination reduction and reinforcement learning explained clearly and in depth.
What a base LLM is and how it is built: Common Crawl and FineWeb data, tokenization, Transformer pre-training and why the base model is not a chatbot yet.