Engineer cuts AI coding tokens by 94% with local code index

Reduced per-query tokens from 45,000 to ~5,000 by sending only relevant context via a local code index. Significantly cut cost and latency in production.