Math Scanner — iPhone PWA: Camera + Google Gemini AI on Cloudflare Workers (React, TypeScript)
iPhone PWA that captures a photo of a math problem and returns the solved answer in seconds. Vision and solving happen in a single Gemini 2.5 Flash call from a Cloudflare Worker — no server, no storage, no polling.
ReactTypeScriptVitePWACloudflare WorkersCloudflare PagesGoogle Gemini APIAIVision OCRServerlessiPhoneGitHub Actions
Published May 26, 2026
Live · math-scanner.pages.dev




Full description
A purpose-built mobile scanner that does exactly one thing well: phone snaps a math problem → AI returns the solved answer.
Flow
- iPhone PWA opens the native camera via
<input capture="environment">. - The captured image is base64-encoded and POSTed to a Cloudflare Worker.
- The Worker forwards the image + a terse prompt to Google Gemini 2.5 Flash.
- The model's answer is returned in the same response — no taskId, no polling.
- Phone renders the answer text. Tap "Scan another" to repeat.
End-to-end latency: ~3-5s for a single problem, ~15-20s for a full page of problems.
Why it's interesting
- One round trip. Earlier iteration used a local daemon + filesystem + Claude Code agent watching for new images. Replaced all of that with a single Cloudflare Worker → Gemini call. The app is now autonomous and serverless in the literal sense.
- Zero infra at rest. No always-on server, no database, no queue. The Worker runs on demand, Cloudflare Pages serves the static PWA, image data lives only in transit.
- Confidence markers in the prompt. The Gemini prompt asks the model to append
?and a parenthetical note to any line where a digit or operator was unclear in the image — so the user sees "did you mean 6 or 8?" rather than a confidently wrong answer. - Truly free to run. Both Cloudflare Workers (100k req/day) and Gemini Flash (1,500 req/day) free tiers cover personal usage forever.
Architecture
iPhone PWA Cloudflare Worker Google Gemini
(Vite + React) ─────▶ (math-scanner-api) ─────▶ (gemini-2.5-flash)
POST serverless, edge vision + reasoning
image single fetch call in one model
◀──── { answer } ────
◀──── { answer } ────
Tech
- React 18, TypeScript, Vite — static PWA, ~146 KB JS gzipped
- Cloudflare Workers —
wranglerdeploy,account_id-scoped secret for the Gemini key - Google Gemini 2.5 Flash — direct REST call, no SDK (Workers-runtime compatible)
- iPhone-native camera input (
<input type="file" capture="environment">) - GitHub Actions CI — typecheck + build for both app and worker
Repo
https://github.com/ohrytskov/math-scanner