I've been building a small API for one boring, expensive problem: turning receipt and invoice images into structured JSON. Here's the reasoning and the schema, and I'd like a sanity check from people who've actually integrated document OCR.
The problem
If you need to pull structured data out of receipts or invoices, the default options are Azure Document Intelligence and Google Document AI. They're good, and they're built for enterprise document pipelines: model training, setup, and pricing around $0.10 to $0.30 per page once you're off the free tier. That's fine at a million pages a month. It's rough if you're a solo dev or a small SaaS parsing a few thousand receipts for an expense feature.
The insight
Vision-language models got good enough that this is now a single prompt, not an ML pipeline. You hand the model an image and a schema, and it reads the receipt the way a person would: vendor, line items, tax, total. No training set, no per-vendor templates, and it handles messy phone-camera photos and multi-page PDFs reasonably well.
The interesting part is the cost. A vision-LLM extraction runs somewhere around $0.001 to $0.002 per page in raw inference. The incumbents charge $0.10 to $0.30. That is a 50x to 100x gap on the same job, and it exists because their pricing is anchored to the old OCR-plus-rules-engine world, not to what a model costs today.
How it works
One endpoint. POST an image or PDF, get typed JSON back.
POST /v1/extract
Content-Type: multipart/form-data
X-API-Key: your_key
file=@receipt.jpg
{
"vendor": "Blue Bottle Coffee",
"date": "2026-08-14",
"currency": "USD",
"line_items": [
{ "description": "Latte", "qty": 2, "unit_price": 5.25, "amount": 10.50 },
{ "description": "Croissant", "qty": 1, "unit_price": 4.00, "amount": 4.00 }
],
"subtotal": 14.50,
"tax": 1.16,
"total": 15.66,
"payment_method": "VISA ****4471"
}
No SDK required, no per-vendor template setup. The goal is that you drop it into an expense tracker, a bookkeeping tool, or an internal ops script and get clean fields out.
The honest part
It's pre-launch. I'm validating that the pain and the price are real before I put serious time into hardening it, so I don't have independent accuracy benchmarks to quote yet, and I won't make one up. What I'm confident about is the cost structure, which is just model pricing versus incumbent pricing.
Planned pricing is around $9/mo for 1,000 pages up to $79/mo for 20,000, with a free tier to test against.
What I'd like to know
If you've integrated Azure or Google's document APIs, or rolled your own with Tesseract or a vision model: what's actually missing from the JSON above that you'd need in production? Confidence scores per field? Bounding boxes? Webhook delivery for async batches? That is the stuff I want to get right before launch.
If it's useful, the waitlist is here: https://ocr-waitlist.alphaai-services.com/?src=devto
Genuine feedback beats a launch-day spike. Thanks for reading.
Top comments (0)