Writing
Notes from production AI products
Engineering notes from production AI products: LLM calls, retries, caching, queues and payments in Talkemi, Fundai and MrPicasso, with the code, tests and trade-offs. Each piece follows one real problem from symptom to fix, with what was tested and what is still not promised.
-
One rebuild, not twenty: what caching a live market dashboard taught me
A Fundai market card made one investor in four wait up to nine seconds. How versioned keys, a Redis lock with a token and a kept copy turned twenty rebuilds into one.
-
Two short transactions around an LLM call
How Talkemi keeps a slow LLM call outside the database lock without losing moderation, failed replies, ordering or retries: two short transactions and an explicit message state.
