Free to try, right now · no signup · real Groq API underneath

Messy PDF in. Clean data out.

This is the same pipeline I'd build for your documents — try it below, then tell me what's different about yours.

Live — invoices + contracts

See how it works under the hood →
Document Type
Invoice

Drop your document here

or browse files

PDF · PNG · JPG — up to 10 MB
262 / 274 fields correct · gpt-oss-20b · 32 test documents (12 adversarial)0 extraction failures in the OCR run · 50 / 50 receipts scoredevery number above links to a committed results file

How it works

From raw PDF to structured data in four steps.

Upload a document, let three agents handle parsing, extraction, and validation, then review any flagged fields and export clean results.

01

Upload

Drop any PDF invoice or contract into the zone below.

02

Extract

Three agents parse, extract and validate every field.

03

Review

Low-confidence fields surface for one-click approval.

04

Export

Download JSON or CSV. Webhook delivery exists in the API, opt-in.

Architecture

Built like production, not a weekend demo.

A queue instead of inline processing. A second LLM call to grade the first. A calibrated judge instead of a self-graded score. These are the decisions that separate a working pipeline from a prompt in a loop.

Parser agent

Detects document type, normalizes the raw content, and prepares the extraction graph.

Extractor + validator

Typed schemas, field-level confidence, and structured outputs keep results operational instead of approximate.

Review-ready output

Low-confidence values route into a review queue so humans only touch what actually needs attention.

Export layer

JSON and CSV export today; a signed, SSRF-guarded webhook exists in the API for anyone wiring this into their own system.

Measured, not marketed

Every figure below comes from a committed results file.

No case study copy. The evaluation harness, the golden dataset, and the raw run logs all live in this repository — click through and check the math yourself.

Live evaluation results

262 / 274
Field checks passed, gpt-oss-20b
32 documents. gpt-oss-120b: 265 / 274. The 3-field gap is within run-to-run noise: the same 20 cases scored 165, then 160 / 170 on two 20b runs.
86.25%
OCR text-layer baseline
Mean over 20 cases, synthetic evaluation, gpt-oss-20b
0.993
Calibrated judge score
termination_clause, GEval mean over 14 cases; the fuzzy matcher passes 5 / 14 (35.7%)
8000 TPM
Free-tier ceiling, handled
Redis token budget, defers not fails

Honest about what's not measured

Cost-per-document isn't reported here because no run has captured per-case token usage yet. A confident-sounding number with nothing behind it is worse than admitting the gap.

Extension path

New document types are a plugin — one file implementing a shared interface, not a rewrite of the pipeline that already works.

Who built this

Pavan — this is a working system, built and measured end to end.

Every layer here — the queue, the confidence gate, the rate-limit handling, the calibrated LLM judge — was a deliberate call, written up in this repository's architecture decisions rather than left implicit. If you're hiring or have a document pipeline that needs the same rigor, I'd like to talk.

Get in touch

The easy part is a clean invoice. This is built for what isn't.

PM

From the desk of Pavan

No forms, no funnel

Pavan here — I built this one to survive that: a gate that catches invoices whose numbers don't add up, a review queue for anything the model isn't confident about, nothing shipping unchecked. So far the edge cases have been European invoices with comma decimals, a PDF with a prompt-injection attempt written into the text, a contract that quietly supersedes an earlier one. Your documents will break it in a different way — tell me how, and I'll tell you honestly whether it's worth building for.

No forms, no sales call. Just write to me directly, at:

pavanmanikanta45@gmail.com

Prefer to read the code first? Here's the source and every architecture decision behind it.

Pavan Manikanta