SafetySignal Desk: Multi-Agent Recall Review

SafetySignal Desk is a multi-agent product-safety recall review room built on Band. When complaints arrive about a snack bar, seven specialist AI agents collaborate through Band as registered participants with distinct roles and identities to prepare a human-reviewed recall escalation decision. In the demo, Complaint Intake extracts structured facts from raw complaints. Pattern Detection finds a same-batch cluster. Label & Ingredient compares the consumer label against the supplier sheet and catches a suspected undeclared peanut allergen. Batch Trace quantifies market exposure. Recall Precedent pulls real records from the openFDA Food Enforcement API. Regulatory Risk scores the case on a transparent 100-point rubric. Customer & Retailer Response drafts safe communications, but every outbound message remains a draft until human approval. Band is the actual collaboration layer, not a notifier. Agents hand off work by mentioning each other, share structured findings as Band messages and events, and build a live evidence chain that the human recall manager can review. The final output is an audit-ready Safety Decision Packet plus the full Band transcript as the audit trail. The system is deliberately hybrid. Counts, scores, batch exposure, and allergen mismatch checks are deterministic Python for auditability. LLMs handle language tasks such as extraction, summarization, reasoning support, and drafting, each with deterministic guardrails. The AI never issues recalls, sends notices, or makes regulatory filings on its own. It prepares the evidence, blocks unsafe actions, and routes the final decision to a human recall manager.