A harassment shield for LinkedIn DMs. A Chrome extension (Plasmo, Manifest V3) hides abusive messages behind a warning, flags a sender once five people have hidden them, and turns their messages into a PDF evidence report. Every message is screened cheapest first: a word list, verdicts already cached, and only then the model.

We built it as a team of four, and it took first place at HackWIE 3.0 and Code Kshetra 2.0.

What I built: the backend (Express, MongoDB, Redis, Gemini, Puppeteer). The API; moderation moved into its own service with failover across deployments, so a spent quota never stops a check; a verdict cache so no message is paid for twice; and evidence reports that survive bad model output. In the extension, the popup and the login hand-off.

Write-up: https://rajveers.com/projects/safire

Code: https://github.com/rajveeerr/Safire

Live: https://safire-five.vercel.app