The problem
Most OCR is good enough for printed paragraphs and useless for everything else. The moment a document has a table, a handwritten signature, a stamp, or a column break, accuracy collapses and someone has to re-key it. For businesses processing thousands of scans a week, that’s a real cost.
What we’re building
Screenshot to Text is a Flask-based SaaS that accepts images and PDFs, runs them through a layered pipeline — Tesseract for the easy text, GPT-4 Vision for the hard stuff (tables, handwriting, mixed layouts) — and returns structured output with per-block confidence. Stripe handles billing on a per-page tiered plan. Cloudflare R2 stores the originals.
The AI angle
Two-pass extraction. The first pass is deterministic OCR; the second pass uses GPT-4 Vision only on regions the first pass scored below a confidence threshold. This keeps cost down (Vision calls are expensive) while raising the floor on accuracy where it matters.
Routing on confidence rather than running Vision over every page is the part that makes this economical at volume. Most of a typical scan — body paragraphs, standard form fields — is exactly what Tesseract is already good at, so paying for a Vision call there would be pure cost with no accuracy gain. Reserving the expensive model for the regions that actually need it (a table with merged cells, a signature, a stamp overlapping text) is what lets the accuracy floor rise without the per-page cost rising with it.
Per-block confidence scoring is also what makes the output usable downstream rather than just accurate on average. A single page-level confidence score tells an integrator nothing about which fields to trust; per-block scores mean a business can auto-accept the high-confidence fields and route only the uncertain ones to a human, instead of treating every extraction as equally reliable or equally suspect.
How it’ll be used
- Bookkeeping firms processing client receipts at scale.
- Healthcare admin teams digitising patient intake forms.
- Legal teams with handwritten case notes that need to be searchable.
Where we are
Core extraction pipeline works. Stripe billing and per-tenant rate limiting are in. Public launch is gated on an SLA we can confidently sell — we’re currently soak-testing against a target of 99.5% per-page accuracy on the standard test set.