Project · Active product
GrimReaper
AI browser personas stress-test web apps, find failure paths, and turn the wreckage into actionable reports.
Product
GrimReaper sends different AI browser personas through a submitted web application to uncover UX friction, broken paths, and failure modes. It classifies what went wrong and produces a useful report instead of stopping at a generic automated scan.
The project began around the GrowthX Hermes AI Agent Buildathon, where it finished in the Top 5, and continued afterward as a serious product experiment.
Architecture
Cloudflare handles the public surface and submission API. Convex stores durable scan state, acts as the job queue, and streams progress to the UI. A long-running Python service on a DigitalOcean VM claims scans, coordinates Hermes manager and persona agents, and writes structured results back to Convex.
The design uses claim, heartbeat, and completion semantics with bounded concurrency and duplicate-processing protection. Screenshots and shareable artifacts stay outside Convex in object storage.
Reliability
Optional services degrade independently: context discovery, text-to-speech, and payments must not block the core scan. Writes are idempotent, browser failure can fall back to lightweight inspection, and multiple runners should never process the same scan.
The model is one component inside a larger system that also has queueing, realtime state, browser execution, external-service fallbacks, and public result generation.