ShiftGuard: AI Shift-Handover Workbench

In high-volume food and beverage retail, store shift handovers happen every 6 to 8 hours under tight time pressure. Outgoing baristas and shift leads communicate operational issues via hurried, multilingual voice notes and shorthand messages (e.g., "tadi grinder macet 20 menit, buang beans 200g, es batu abis"). Because standard POS and ERP systems only capture clean transactional data, discrepancies between frontline notes and back-office records slip through unnoticed, resulting in phantom inventory, unlogged waste, and equipment stalls during peak morning rushes. ShiftGuard is a commercial shift-handover reconciliation workbench built for the Reinvent Commerce track. It combines Google Gemma 3-27B-IT running on AMD Instinct GPUs (via the AMD Token Factory API) with a deterministic 5-rule verification engine to turn messy shift notes into verified, actionable store handovers. How ShiftGuard Works: 1. Multilingual Claim Extraction (Google Gemma 3-27B-IT on AMD): Parses colloquial Indonesian and English shift notes into structured operational claims while enforcing an anti-hallucination exact substring quote gate. 2. Deterministic 5-Rule Verification Engine: Cross-checks every extracted claim against 10 store ground-truth records (POS sales exports, void logs, cash counts, espresso cycle counters, cleaning logs, incident books, inventory ledgers, stock sheets, and supplier purchase orders). 3. 3-Axis Epistemic Triage Ledger: Categorizes findings across Severity, Evidence Status (Supported, Contradicted, Not independently verified), and Workflow Status. Gemma generates specific technician and vendor mitigation actions that supervisors can apply in one click. 4. 30-Second Incoming Briefing & Audit Trail: Unlocks shift sign-off only after critical blockers are resolved, synthesizing a 30-second executive briefing for the incoming shift lead alongside an append-only audit log.