Longer-form writing about the systems I ship: multi-agent retrieval, real-time voice, and the unglamorous infrastructure decisions that decide whether any of it survives production.
A word like "cat" can be a pet, a predator, or something that falls off a mountain — so how does one point in a vector space hold all of that? Embeddings make far more sense as a learned map of meaning than as a list of properties. Here's what a dimension actually is, and why context does most of the work.
AI models have no memory, so you resend the whole conversation on every message and the model re-reads it from scratch. Prompt caching is the fix — reuse the computation for the unchanged prefix and cut API costs 50–90%. Here's what's actually being cached, and the attention math behind why it works.