I Wanted AI To Do More
Than Read Documents

As an AI systems engineer, I’ve watched models become remarkably good at reading documents.

They can recognize text, understand nested layouts, extract tables, and turn raw pixels or PDFs into structured data.

But building something useful for a business requires more than extraction.

An invoice doesn't become processed just because a model can read it. Someone still needs to check the numbers, identify duplicates, resolve inconsistencies, approve the record, and make sure the original document is accounted for.

I saw the gap between AI reading documents and AI producing data businesses can actually trust. So I set out to bridge it.

Financial documents carry numbers that businesses depend on. A misplaced decimal, incorrect tax amount, or duplicate invoice can create problems long after the document has been processed.

Yet many automation tools focus only on getting information out of a document, leaving verification and operational responsibility to someone else.

A plausible answer isn't good enough for financial data.

Instead of building another wrapper that simply returns extracted fields, I set out to build a system that handles the workflow around them.

That idea became Recutr.

I built Recutr as a document operations pipeline that brings document intake, extraction, validation, review, and output into one connected workflow.

Invoices and purchase orders arrive through shared inboxes or direct uploads. A combination of PDF parsing, OCR, and vision AI extracts the relevant information, and the system classifies each document before it is filed.

Today that workflow runs through five agents: Sift classifies, Scribe drafts, Proof checks, Match reconciles purchase orders, and Trace keeps the record.

Sift

Scribe

Proof

Match

Trace

The system then checks totals, line items, dates, amounts, and potential duplicates. Each extracted field carries confidence information and source evidence, making it clear where the data came from.

Records that need attention are routed for human review. Financial approvals remain in human hands.

Approved invoices export to your ledger with their original documents linked. Purchase orders keep their matched invoices and billing balances in one view.

I believe automation should never require businesses to surrender control of their data or blindly trust an AI-generated result.

I designed Recutr around three core principles:

Recutr isn't designed to be another chat window sitting beside your accounting software.

It's an operational system built to take repetitive invoice processing off your team's hands while keeping important decisions visible and accountable.

The value is time returned. Fewer minutes keying each invoice, less time chasing POs and approvals, and a clean trail at month-end.

From incoming emails to approved financial records, every stage has an auditable place in the workflow.

I'm starting with invoice automation and building toward a broader vision of intelligent document operations.

Built by one engineer and a small team of machines:

I am building in the open, one pipeline at a time: