A small French API quietly took on the dullest job in the office - copying numbers off a PDF - and turned it into a developer product the internet actually thanks.
Picture the inbox. Forty-three invoices in a single ZIP. A receipt photographed at a slight angle with thumb included. A passport scan from a contractor in Lagos. A bank statement, five pages, in PDF, with three different fonts. Someone, somewhere, is about to retype all of this into a form. That someone is the reason Mindee exists - and the reason they will not have to do it for much longer.
Mindee is a Paris- and San Francisco-based AI company that does one thing with quiet conviction: it turns documents into clean, structured JSON. Not summaries. Not vibes. Structured fields, with confidence scores, available through an API call that takes about as long as a sneeze.
Founded in 2018 by four ex-AI executives - Jonathan Grandperrin, Olivier Rey, Victor Briançon-Marjollet and Mohamed Biaz - the company spent two years in stealth, built quietly, and then emerged in October 2021 with $14M led by GGV Capital, Alven, Serena Capital and Bpifrance, with cameo checks from Algolia's Nicolas Dessaigne and Datadog's Alexis Lê-Quôc. A few months later they joined Y Combinator's Winter 2021 batch - one of the few French AI companies to make the journey.
Mindee ships pre-trained endpoints for the documents the world keeps generating, plus a no-training-required builder for the documents the world makes up daily. The catalogue reads like a list of every form you have ever cursed at.
Line items, totals, taxes, vendor info - parsed with item-level itemization. Used by expense management platforms across Europe.
Driver's licenses, passports, healthcare cards and international IDs across dozens of countries.
Bank statements, payslips, energy bills. The financial paperwork no one volunteers to type.
Training-free document AI - bring any contract, label or form, get a tailored parser in minutes, not months.
Break apart multi-page uploads, sort by type, isolate multiple documents on a single scan.
Pipe models together into a single API call - the JSON arrives already opinionated.
Three rounds, one San Francisco zip code, and a cap table that runs from boutique French funds to GGV in Menlo Park.
Investors include GGV Capital, Alven, Serena Capital, Bpifrance Digital Venture, Algolia co-founder Nicolas Dessaigne and Datadog co-founder Alexis Lê-Quôc.
Former CTO at TankYou and ECTOR. Runs the San Francisco beachhead.
Engineering lead behind the OCR stack and the proprietary LLMs in beta.
Runs operations across Paris, San Francisco and Montreal.
Science lead - the brain behind the document models.
Mindee's logo wall reads like a tour of European fintech and HR-tech. The same API powers expense reports for Spendesk, payroll documents for PayFit, and procurement workflows at Lucca - plus a long tail of developers shipping side projects on the free tier.
Most of the AI press cycle in 2026 is still about chatbots. Mindee's bet is the opposite: the next decade of enterprise value sits in the dull middle - the layer between the document and the database. That layer has been duct-taped together for years with regex, Mechanical Turk, and increasingly creative interns.
Their pitch is simple. Data scientists spend up to 39% of their time preparing data. docTI deletes most of that. A product manager can build, test and integrate a custom document parser in a single afternoon. Multiply that across every fintech, insurance carrier, logistics company and HR platform that touches paper - and you have an infrastructure category hiding in plain sight.
Mindee is not loud about it. It does not need to be. The customers are happy. The reviews are good. The API does what it says, and quietly, line by line, makes the world's filing cabinets readable.
Forty-three invoices. The receipt with the thumb. The five-page bank statement. The Lagos passport. None of it is being retyped anymore. A single API call - probably running in a containerized backend somewhere in Spendesk or PayFit - returned a JSON object with vendor names, line items, totals, dates, and a confidence score for every field.
The person who used to retype is now reviewing exceptions. The accounting close happens in days instead of weeks. The intern got reassigned to something that requires a pulse. And somewhere in Paris, an API silently logged the request, charged a fraction of a cent, and moved on to the next forty-three invoices.
That is Mindee's contribution to the workday: not louder, just less typing. The boring revolution is here, and it is filed under application/json.