AI document processing

AI document processing and data extraction

Invoices, PDFs, forms, emails and scans hold data your team types into other systems by hand. With AI document processing, that data comes out as clean, structured records that go straight into your software. I build it for companies without a developer of their own, and for teams that need someone experienced.

I'm all in on AI, so I move fast. I stay careful about accuracy. Wrong data in your system is expensive to clean up.

I usually reply within one working day.

Vincent Schmalbach

What intelligent document processing covers

From the incoming document to a checked record in your system.

PDF data extraction

Line items, totals, dates, names and addresses from PDFs, into fields with a fixed format. The model has to follow a schema, and each result is validated.

Invoices and receipts

Supplier, amounts, tax and due dates read automatically and passed to your accounting or ERP system.

Emails and forms

Orders, requests and inquiries from emails or web forms turned into records, sorted by type and sent to the right person.

AI OCR for scans

Scanned or photographed documents read by models that also understand the layout. Useful when the old OCR tool gets tables wrong.

Checks and review

Rules that catch impossible values, totals that don't add up and missing fields. Unclear cases go to a person with the document next to the extracted data.

Into your software

The records go into your database, CRM, ERP or accounting tool through its API, with a log of what came from which document.

How I work

You don't need to have everything figured out before we talk. An idea, access to the repo, a screen recording or a list of problems is enough. I look at what is there, tell you what I would do and what it costs, and then I build it.

I use Claude Code, Codex and my own agent workflows for reading code, planning changes, writing tests, refactoring and debugging. It is not vibe coding. It is senior engineering with a much faster loop.

AI-first, with checks on every field

Claude Code, Codex and my own agent workflows help me build extraction fast. I test it on your own documents and add checks that flag results you can't trust.

I stay in control

AI writes a lot of my code. I still make every technical decision, read every change before it ships, and test it. Fast, but meticulous: the speed is only worth something if what goes live is right.

The way you want to work

Some clients want to look at every change before it goes live, others want me to deploy. Some want a weekly call, most prefer written updates. I work well asynchronously, and I still ask the important questions when they come up.

I think like an owner

I still run my own online products. So when I build or fix yours, I think about traffic, conversion, revenue and support as much as about the code, and I tell you when something is not worth building.

Examples of my work

Projects where I turned messy input into structured data.

Financial workflow SaaS

Market data in, checked reports out

Users define workflows over market data from several APIs. The product normalises the data and writes AI reports, with validation, credits and Stripe billing inside the product.

Car marketplace

Dealer stock turned into clean listings

Stock from European dealerships imported, enriched with manufacturer data and published as listings on the marketplace and on other car marketplaces.

Outreach platform

Contact details found in scraped pages

A Laravel platform on 30+ servers that reads scraped web pages, finds link opportunities and the contact details of site owners, and stores them as records for automated campaigns.

What clients say

"What stood out about working with Vincent was how methodical and efficient he was. He started by really digging into our requirements, asking clarifying questions until he had a solid understanding."

Philipp Toepelmann, Digital Investments GmbH

"He also turned customer discussions into clear technical requirements and documented his work so the rest of the team could build on it."

Yash Chandra, Academy of Mine

"Vincent was easy to work with and produced high-quality code. Would recommend."

Dustin Fraker, QuickerNotes, LLC

Questions

How accurate is AI data extraction?

It depends on your documents. Before anything goes live, you see the results on a sample of your own documents. Fields that fail the checks go to a person.

Is this better than classic OCR?

For clean, fixed forms, classic OCR can be enough. For changing layouts, tables and handwriting, current AI models read much better.

What happens with our documents?

They are processed inside your system and sent to the model provider only for the request. I explain what goes where, and we pick providers whose terms fit your requirements.

Can it read invoices in several languages?

Yes. The models read German, English and most other European languages, and the output uses the same fields either way.

Can you connect it to our accounting or ERP system?

Yes, if the system has an API or an import. The extracted data goes in with a log of which document it came from.

How do you charge?

Fixed price for a job with a clear outcome, a flat monthly rate for ongoing work, or hourly. Describe what you need and I'll suggest one.

Get a quote

Which documents you get, roughly how many, and where the data should end up. A few sentences are enough. I'll reply with how I would build it and what it would cost.

If you would rather talk first, book a 30-minute call or email contact@vincentschmalbach.com.

I usually reply within one working day.