How to run continuous, AI-powered security audits on your codebase — routed through a local proxy that picks the cheapest viable model for each file. Most security scanners feel like spam filters from 2005. They flag every eval() , every string concatenation that looks vaguely SQL-ish, and every hard-coded constant longer than ten characters. The signal-to-noise ratio is so low that teams either stop running them or learn to ignore the output entirely. The newer generation of scanners uses LLMs to read the code the way a human reviewer would — understanding intent before flagging pattern . Deepsec, from Vercel, is one of the most usable tools in this category. The problem is that running it across a real codebase, file by file, with a frontier model, gets expensive quickly.…