Recalls are common, and especially so for medical devices: 1,059 device recalls in 2024, a four-year high (Sedgwick).
SafeCart remembers a shopper's preferences, wishlist and owned items. When a wishlist item goes on sale and fits their rules, it buys it. Above the $150 auto-buy limit it asks for approval and runs an identity check first.
Before every purchase it checks for recalls. It never auto-buys when the result is unknown. A recalled item is skipped and the next safe fit is bought. When a new recall matches something the shopper owns, it files a refund claim and buys a safe replacement. Every step lands in an audit trail with cited sources.
How we used each tool:
- SingleStore: catalog, prices, wishlist, owned items, recalls and the audit ledger; alternative search.
- Senso: knowledge base of recall notices and store policy; a fact-check gates every buy.
- Superlinked SIE: flags injected instructions in untrusted listing text, extracts brand, model and UPC, reranks alternatives by preference.
- Statsig: purchase policy config and a recall_guard gate we flip live to show a naive agent for contrast.
- Amplitude: tracks agent outcomes (sales, safety checks, buys, recall alerts).
- Persona: identity check over the limit.
- Claude: writes each decision rationale and the refund letter; buy/skip stays rule-based.
Recall data is real (CPSC API; FDA device recalls are next). Payments are Stripe test mode. The shopper and store are simulated.