Home / AI document processing
Invoices, applications, contracts and incoming requests get read, checked against your rules and prepared for a decision. A person approves, and the record keeps who decided what.
Use cases
Invoice numbers, amounts, names and dates move from a PDF or an email into a spreadsheet by hand. A model can read the file; the value is in what happens after that — validation, a queue, an approval and a record.
Extraction is confident even when it is wrong. I build the part that catches it: rules that flag what does not add up, a review screen for anything uncertain, and a full history of what was changed and by whom.
Delivery
Each extracted field keeps a link to the place in the original document it came from, so a reviewer can verify it in a second instead of opening the file.
Totals, duplicates, missing fields, amounts over a limit, an unknown supplier — checks run in code, and anything that fails goes to a person.
A real interface with roles and statuses. Approve, correct or reject, with the document next to the decision.
Every run, every correction and every approval is stored. When someone asks in March why an invoice was paid in January, the answer is in the record.
Example
A private archive built over seven Telegram groups. It imports the history, transcribes voice and video notes, and writes a topic and a short summary for each lesson day. The original messages and materials stay attached, so the summary is a way in rather than a replacement.
React · TypeScript · FastAPI · SQLite · Telegram API · Azure Speech · OpenRouter · Docker
Pricing
Every project starts with a fixed-price process audit: €250, three to five days, credited toward the build. Processing with a review interface then costs €2,500–6,000 and takes 4–8 weeks.
FAQ
Whichever fits the document and the budget, and it is a replaceable part of the system rather than the system itself. Structured layouts often need no model at all — a parser is cheaper and does not hallucinate.
It gets caught by a rule or by the reviewer, corrected in the interface, and the correction is stored. That is the point of keeping a person in the loop: mistakes are expected and cheap to fix, instead of silently written to your database.
Only the part that needs to, and only if you agree to it. The database, the files and the audit trail stay on infrastructure registered to you, in the EU by default.
Yes. Backfilling an existing archive is usually the first run — it is also the cheapest way to find out how accurate the extraction is on your real documents before anyone depends on it.
Related
Contact
Tell me what your team still does by hand, where the information comes from and what should happen next.