Why
Why I built it
Dumping dozens of page photos into an AI at once makes missing pages, scrambled order, and OCR errors hard to catch. Pages need ordering upfront and side-by-side proofing before relying on the text.
What I built
What I built
- Batch selection, drag-and-drop ordering, zoom inspection, and page-by-page transcription preserving page numbers and source filenames.
- Switchable Codex CLI and OpenAI-compatible vision backends, showing per-page progress and actual token usage.
- Integrated from a standalone PWA into a unified scanner where users verify OCR output against the original image before export.
Role & AI
How I split the work with AI
Driven by my own paper backlog, I designed the phone-capture-to-desktop-proofing flow and iterated on implementation with coding agents.
Trade-offs
What testing made me change
- Moved from an early Android prototype to a shared Web/PWA interface to avoid dual maintenance, while keeping batch gallery selection separate from single-shot camera capture.
- The local PWA handles ordering, inspection, and proofing, while OCR calls switch flexibly between Codex CLI and OpenAI-compatible vision models.
Outcome
Outcome
Evolved from a standalone photo-to-Markdown utility into an integrated single-service scanning and proofing workspace.