Achyrix: AMD Gemma Restaurant Intelligence

Achyrix is an approval-gated restaurant operations copilot. Restaurant teams make prep, purchasing, and menu-availability decisions using fragmented signals such as recent sales, forecast demand, on-hand inventory, safety stock, reservations, catering orders, supplier delays, open purchase orders, and refunds. A generic chat assistant may overlook these interactions and produce a plausible but unsafe recommendation. Achyrix uses Google Gemma hosted on AMD infrastructure as its primary reasoning layer. It sends a compact, redacted operational snapshot to an OpenAI-compatible Gemma endpoint and requests an evidence-backed, structured recommendation. Before anything reaches the user, the application validates the response deterministically by checking tenant, outlet, item, and date scope; safe prep and stock limits; known-cost bounds; registered tool permissions; and manager-approval requirements. The model cannot silently alter prep, procurement, or menu availability. The project is containerized and includes a controlled fallback path. If the model cannot respond safely within its request budget, Achyrix returns deterministic guidance instead of inventing an answer. We built a 13-scenario evaluation set covering waste, low stock, purchase orders, supplier delays, promotion spikes, catering, refunds, unavailable items, cross-outlet imbalance, conflicting signals, missing data, malformed data, and normal no-action cases. This combines practical AI assistance with the trust and guardrails restaurant operators need.