ShelfWise - Agentic AI Store Operations

A supermarket runs on hundreds of interlocking decisions a day - most made late, on gut feel. Yoghurt expires because nobody ran the markdown math. A shortage gets answered with "transfer some stock" - which branch, DC, or supplier, at what lead time? Till prices drift and silently eat margin. A fridge failure starts a countdown on thousands in stock. Retail loses its profit in a thousand small decisions made without the math. ShelfWise makes those decisions with the numbers in hand - and never acts without a human signature. Every decision runs through a Critic/Executive agent pair powered by google/gemma-4-E4B-it on an AMD Instinct MI300X GPU via vLLM on ROCm, with native tool calling. Agents reason through a bounded loop over 11 read-only tools - stock, demand forecast, expiry risk, reorder policy, supplier ranking, cold-chain risk, price integrity, markdown simulation, and a multi-source sourcing optimizer ranking branches, DCs, and suppliers by availability, distance, and lead time. A grounding guarantee is enforced in code: an answer citing a figure no tool produced is rejected and re-run - never shipped. Every recommendation lands as a pending decision requiring human approval before any write-back - full audit trail, tenant isolation, and a learning loop that adjusts thresholds from real outcomes. Chat is the front door: one question triggers multiple live tool calls and returns a structured report. Beneath it, each store gets a digital twin - an evolving virtual replica where agents train and prove themselves before touching the real shop floor. Proof, not promises: a 15-minute live soak on the MI300X endpoint - 333/333 chat calls model-backed, zero fallbacks, 4,618 decisions, zero ID collisions, 2,990 HITL cycles, zero mismatches, answers < 6s. 415 tests pass; CI verifies a 165-capability manifest. Retail AI demos well and governs badly. ShelfWise inverts that: agents that show their math, a human on every action, receipts for every claim.