A working notebook. What I built, broke, read, and thought about — short entries, written close to the moment. Not a log. Just the notebook.
Published tlme to npm — a terminal AI assistant that streams answers inline. Ask a question and it talks to Claude, OpenAI, Grok, Gemini, or OpenRouter behind one interface, with zero npm dependencies—just native fetch and node:*. The fun part was hand-parsing five different SSE dialects and building an arrow-key setup wizard without a single UI library.
Wrapped up an AI-first CRM for pharma field reps — a read-only interaction form driven entirely by a LangGraph agent. Reps describe a visit in plain English and the assistant extracts, edits, and saves every field. Eight tools, full CRUD to Postgres, running on a Groq LLM.
Spent the day getting a LangGraph agent to route cleanly across eight tools. The small models kept looping instead of stopping; gpt-oss-20b turned out to be the sweet spot—reliable routing without the 70B price tag. The real trick was a strict system prompt and an edit tool that merges corrections instead of overwriting the whole form.
Spent a few hours understanding FastAPI beyond CRUD examples—dependency injection, request validation with Pydantic, async endpoints, middleware, and project structure. Feels like a much cleaner fit for AI backends than Flask.
Read through how Redis works under the hood and where it actually makes sense. Looked into caching API responses, session storage, rate limiting, pub/sub, and queues. It's much more than 'just a cache.'
Sketched a cleaner flow for MedCap. The goal isn't adding features—it's reducing friction. Simplified the upload process, reorganized result sections, and explored ways to make AI-generated insights feel easier to trust.