Vera

Problem

Solo builders and teams build features nobody asked for. The evidence — real demand, or a competitor already shipping it — sits in Reddit threads, X posts, and changelogs nobody reads. Nobody is doing customer interviews at 2am, but nobody wants to burn hours on something with no demand either. So the ticket gets built on a hunch.

Role

Solo build, a few hours. @mention Vera on a ticket and it researches Reddit, X, and competitor changelogs, then posts a verdict: Build / Investigate / Skip, with demand, urgency, and real quotes. The model picks evidence by ID; code enforces the verdict, so quotes can't be fabricated. Standing monitors flag competitor changes relevant to open issues.

Outcome

1st place at Cognition's Devin × LvlUp Ventures Build Night. Live at askvera.dev. Real runs pull genuine signal fast — one validation gathered 34 signals across 26 pages in 68 seconds; median runs land around 90 seconds. The call that made it work: don't let the model write the verdict free-form. It classifies and picks evidence by ID, and deterministic code assembles the comment — which is why every quote in a Vera verdict is real. Next: richer monitor coverage and letting teams tune the verdict policy per project.

Stack

• Next.js
• TypeScript
• trigger.dev v4
• Linear agent API
• Claude API
• Context.dev
• twitterapi.io
• Vitest
• Vercel