N° 01 SafetyPulse
A safety regression gate for LLM fine-tuning
14.3% → 25.2% on 119 HarmBench test promptsASR rise from benign Alpaca tuning
Mandate lets a team set spending rules that an AI shopping agent must operate within. The LLM only proposes products and quantities; a deterministic TypeScript policy engine decides ALLOW, NEEDS_APPROVAL or DENY, and every decision is written to an append-only audit log. Built for the PayPal AI Hackathon 2026 (sandbox only).
Giving an AI agent a payment method is currently all-or-nothing, exposing teams to overspend, price drift, prompt injection, structuring and lack of accountability.
Treasurer defines a mandate: budget, per-item cap, categories, approval threshold
LLM parses plain-language requests and ranks catalog items, never handling prices
Deterministic policy engine recomputes prices and returns ALLOW / NEEDS_APPROVAL / DENY
Payments via PayPal Orders v2 sandbox; human approvals via Zapier
Dashboard (AG Grid) over an append-only audit log
N° 01 A safety regression gate for LLM fine-tuning
14.3% → 25.2% on 119 HarmBench test promptsASR rise from benign Alpaca tuning