Permanent QR code hosting for VisionLink: every scan of a code printed on a business card or vehicle goes through it, and companies manage their codes from a dashboard or a B2B API.
Lead developer and the only engineer on the rebuild. I designed and built the Go scan service, the database and the dashboard, and ran the cutover from v1. We planned it together, and VisionLink's co-founder handled the business side.
Go with Fiber and pgx, Next.js 16, React 19, TypeScript, Tailwind CSS 4, TanStack Query, zod, Supabase (Postgres with row-level security, Auth), pgTAP, GitHub Actions. The scan service runs on Railway and the dashboard on Vercel.
flowchart LR
L[Printed v1 URL] --> R[Next.js legacy route]
S[Short URL] --> G[Go scan service]
R -->|302| G
G -->|302, no-store| T[Destination]
G -.->|batched scan events| DB[(Postgres)]
D[Next.js dashboard] <--> DB
A[B2B API clients] --> G
The Go service is the only place a code is resolved or a scan recorded. It always answers with a 302 and no-store, never a 301, so a destination can change after printing. The dashboard reads through row-level security and writes through database functions, with pgTAP tests in CI.
Printed codes can't be recalled. v1 URLs are printed on metal cards and vehicles, so they have to work forever. I rejected porting v1's redirect, create and stats endpoints into v2, because that would carry over v1's quirks and split code lookup across two services. Instead a thin legacy route forwards every printed URL to the Go service. Before the swap I audited every v1 code against the new code rules and followed each printed URL through both hops. Every printed code resolved at cutover.
A scan must never fail. I rejected rate-limiting scans with a 429 the way the API is limited, because the person scanning has no retry button. Scan analytics go into a bounded in-memory queue that writes in batches, so a slow database never holds up a redirect. When the queue is full or one source floods it, the analytics event is dropped and the scan still redirects.
Live. Production moved to v2 in September 2026. p95 redirect latency dropped from about 4 s to 312 ms.
Source is private. Full write-up: https://mathysbasson.vercel.app/work/visionlink-qr-v2