REPOMIND v3 - on-prem AI coding agent on AMD

REPOMIND v3 is an on-premises, air-gapped AI coding agent - plus an LLM cost-router - built 100% on AMD. The problem is two-sided. Regulated enterprises (banks, defense, healthcare) are legally barred from cloud AI coding tools - Samsung, JPMorgan and Apple have all banned them - so their engineers have nothing. Meanwhile everyone else is bleeding on token cost: Uber burned its 2026 AI-coding budget in four months, and Microsoft pulled its internal Claude Code licenses over cost. One engine serves both. On-prem, REPOMIND runs open-weight models on the customer's own AMD hardware (Radeon → Instinct), so code never leaves the perimeter and inference is $0 per-token. The same engine is a cost-router that cuts existing Claude/OpenAI/Gemini bills 40–70% and migrates the cheap majority of traffic to AMD-hosted open-weight. It is 100% AMD with no closed models: a Gemma-class local router delegates to AMD-Instinct-powered Fireworks for deep reasoning. A measured cost layer cuts tokens via AST skeletonization (−89%), prompt caching (−90%), and a confidence cascade that skips the expensive model when the cheap one is sure. Every agent action is written to an Ed25519-signed, hash-chained AI Flight Recorder - turning EU-AI-Act compliance into automatic evidence. REPOMIND won 1st Place (AI Agents) at AMD ACT I, verified on a real MI300X: a 256K-token repo served for $4.12. It ships today - docker compose up, 38 passing tests, and an offline demo needing no GPU or keys. This is the cost + trust layer for on-prem enterprise AI on AMD.