What I learned building an internal Q&A assistant where the AI answers from your documents — and my code, not the AI, decides whether the answer counts.
Eight design rules from shipping a browser-driving LLM script — how the deterministic scaffolding around the model becomes the whole game.
How we kept token bills bounded while translating hundreds of pages of Japanese business docs through Claude.
How we kept WebRTC video sessions alive across deploys on a real-time consultation platform during the COVID demand spike.