Customers panic
When recalls hit the news, even safe customers throw away clean food and demand refunds. Stores lose money and trust.
AI RECALL OPERATING SYSTEM
Trace recalls. Notify the affected. Reassure the safe.
Pheromone is the first multi-agent recall system that proactively reaches both affected and safe customers — backed by supply chain provenance proof.
AMD
Running Qwen3-32B on AMD Instinct MI300X via vLLM + ROCm 7
The problem
When recalls hit, systems are built for containment — not for trust.
Stores can’t pinpoint
The data exists — receipts, loyalty cards, transactions — but stores don’t have the tools to use it under pressure.
Affected don’t find out
84% of grocery chains have no purchase-level notification policy. Affected customers learn from the news, days late.
3,000 deaths a year from foodborne illness. This isn’t a tech gap — it’s a tools gap.
Five agents. Each solves a piece of the chaos.
A calm pipeline: structured intake, defensible trace, probabilistic matching, operations, and comms.
Intake
Parses messy recall notices into structured JSON
Trace
Walks the supply chain backward through recursive SQL
Match
Computes per-transaction probability of affectation
Ops
Generates store-side actions (pulls, blocks, tasks)
Comms
Drafts customer messages in the right tier and language
How probabilistic provenance actually works.
In a real store, lots get mixed. Pheromone doesn’t pretend to know. It computes probability.
Customer A
Purchase at 8:05am. Shelf composition: 3 units (1 affected, 2 safe). Probability of affected pick: 33%.
Tier: POSSIBLE AFFECTED. We have identity (card) — direct message sent.
Customer B
Purchase at 11:30am. Cash payment, identity unknown. Shelf state at time: 2 units (1 affected, 1 safe). Probability: 50%.
Tier: POSSIBLE AFFECTED. Can’t message directly — issue a specific public notice with a time window.
Customer C
Purchase at 2:14pm. Shelf state: 1 unit, the safe one. Probability of affected pick: 0%.
Tier: CONFIRMED UNAFFECTED. Reassurance sent with provenance proof.
This is the part no other recall system does: specificity of public notices, reassurance with proof, probability instead of guessing.
Existing apps notify only the affected. Pheromone reaches both.
Time-bound. Defensible. Proactive.
AFFECTED CUSTOMER
Confirmed Affected · drafted in Spanish (or the customer’s preferred language)
Acción urgente: retiro RG-4429 (high)
Su compra está confirmadamente afectada.
Peligro: Salmonella. Severidad: alta.
No consuma el producto. Devuélvalo para reembolso.
Nota de riesgo: mayor riesgo para bebés/niños pequeños y personas vulnerables.
Prueba de procedencia: su compra proviene de la Planta 2 según inventario al momento de la venta.
As of 2026-05-10T04:22:53Z.
REASSURED CUSTOMER
Confirmed Unaffected · the killer feature
Update on recall RG-4429
Good news: based on our records, your recent purchase came from Plant 1, not the affected Plant 2.
We’ve verified this against shipment records. No action needed.
We will notify you immediately if recall scope changes.
Provenance check completed at 2026-05-10T04:22:53Z.
Validation + stack
67/67
tests passing
5
specialized agents
6
confidence tiers
25
real OpenFDA fixtures
Qwen3-32B
vLLM 0.17.1
ROCm 7.0
AMD Instinct MI300X
192GB HBM3
LangGraph
Postgres
Next.js