I was paying $200/month in wasted AI tokens. So I built a Rust context optimizer.

A developer discovered that over 60% of their AI coding tool's token usage was wasted on duplicate and irrelevant code fragments due to naive Top-K similarity-based context selection. They built Entroly, an open-source Rust-based context optimizer that sits between the IDE and LLM, using KKT-optimal knapsack bisection for diverse fragment selection, SimHash deduplication, and an online RL algorithm (PRISM) that learns which context is actually useful. On a 50K LOC monorepo, it reduced token spending by 78% while adding under 10ms latency per request.