Pheromone

When the FDA recalls a food product in the U.S., no law guarantees direct notice to the people who bought it. Recall notices usually sit on agency websites, press releases, or store signs. Meanwhile, 109M units were recalled in the first nine months of 2025, and foodborne illness affects 48M Americans annually. Pheromone is a multi-agent AI recall operating system that closes this notification gap. When a recall is issued by FDA, USDA, or a supplier, Pheromone parses the messy notice, traces the supply chain backward, identifies affected products by store and time window, estimates per-transaction risk without checkout lot codes, and drafts the right message for each customer across six confidence tiers. Its key differentiator is Reassurance Notifications. Existing tools notify only affected customers. But when a recall hits the news, unaffected customers panic too: they throw away safe food, demand refunds, and lose trust. Pheromone also tells clean customers why they are safe: “Your Salsa-Verde purchase came from Plant 1, not affected Plant 2. As of this UTC timestamp, your purchase is clean. We’ll notify you if scope changes.” Architecture: five agents plus deterministic systems code. Intake and Comms use Qwen3-32B; Trace uses recursive Postgres CTEs; Match replays inventory composition to compute pallet-level probability; Ops creates store tasks and POS blocks; Compliance Logger records every state transition. LangGraph orchestrates agents with Postgres-backed checkpointing. Pheromone runs on one AMD Instinct MI300X with vLLM and ROCm. Its 192GB HBM3 memory lets agents share live context without paging. The system passed 67/67 tests, including real Qwen3-32B validation on OpenFDA fixtures and multi-tier scenarios. Target users include independent grocers, regional chains, and large retailers preparing for the FDA Food Traceability Rule. Pheromone enables surgical POS blocks and customer notifications instead of broad inventory pulls and vague press releases.