VisionAir

VisionAir is a live LLM-integrated web app built on Next.js and TypeScript and deployed on Vercel. It takes an ambiguous idea, the stage where most projects stall, and works it into a structured blueprint and a concrete starting point. Its Build Mode turns that blueprint into a complete, ready-to-build context pack for AI coding agents.

The back end is Next.js API routes with typed schemas and request validation, scrypt authentication, HMAC-signed httpOnly sessions, CSRF and origin checks, and server-only secret management, so no key ever reaches the client. The Claude integration routes across two model tiers with a deterministic fallback when a call fails, rather than leaving the app at the mercy of one model call.

The security work is the part I would point at first. It deploys on Vercel with per-PR preview environments and production-scoped secrets, hardened with CSP and HSTS headers and Trivy and Dependabot scanning in CI. It also carries a SECURITY.md threat model covering prompt injection, rate limits, and token and cost caps, which doubles as the pre-commit review checklist. That is the intersection this project exists to prove: full-stack, LLM application, and application security in one build I own.

Both links above are live: the deployed app and the public source. Open either to verify, because the whole point of this site is that you can.