MixMind — A Plant-Private Concrete-Mix AI on AMD

Concrete is the most-consumed material on Earth and roughly 8% of global CO2. The know-how to fix a mix sits in corporate labs most plants can't afford — and no plant will upload its recipes to a third-party AI, because the mix design IS the business. MixMind is a machine operator's answer: a plant-private AI on the plant's own AMD hardware, so the recipe never leaves the building. What it does: 1) Floor Copilot — answers real shop-floor questions from the plant's own notes, cites the note id (e.g. [KB-13]), and refuses when the notes don't cover it. 100% grounded citations on a 9-case held-out set. 2) Mix Committee — four AI specialists (Standard, QC, Cost, Carbon) debate a recipe change, grounded in a strength model trained on 1,030 lab tests (UCI/Yeh 1998, held-out R2 = 0.910) plus deterministic cost/CO2 math. The committee REJECTED the cheapest, greenest mix (45% slag, -40% CO2): predicted 2-day strength 8.0 MPa fails the demould spec. Judgment, not a yes-man. The approved mix cuts CO2 17% with strength held. 3) Mix Optimizer — drag cement/slag sliders and watch the verdict flip in real time. The AMD + Gemma stack: we LoRA fine-tuned Gemma 4 12B on a plant instruction set on an AMD Instinct MI300X (loss 3.03 -> 0.001, 287 MB adapter), judged blind by Fireworks AI. The 12B size is the point: it fits a ~$4k 48GB AMD Radeon card a small plant can actually buy. Everyone proves AMD goes big — we prove it goes small enough to afford. The demo video includes UNEDITED live inference on the MI300X, streaming cited answers token by token. The business: a mid-size plant spends $1.0-1.5M/yr on cement; committee-screened substitution recovers $50K-250K per plant-year. MixMind at $2K/plant/mo pays back in month one, across ~3,000 North American precast plants. Honest boundary: we predict directions, not certified mixes. Evidence logs (rocm-smi, training, eval) are in the repo.