@felixdittrich92OCR and Document AI in Leipzig: Felix, lead ML engineer at t2k
I'm Felix, a lead ML engineer at t2k in Leipzig. I work on OCR, speech recognition and information extraction, mostly in Python. I'm a maintainer of docTR, an open-source library for document text recognition, and of OnnxTR, an ONNX pipeline wrapper for it. I also maintain an OnnxTR OCR plugin for Docling and a tool for generating synthetic text recognition datasets. On the AI-assisted side, my daily tools are Claude, Claude Code and GitHub Copilot.
Right now I'm learning Graph Neural Networks and Rust, and I'm curious how they might fit into document pipelines. I'm also interested in how others are using coding assistants on long-lived open-source codebases.
What I'd love from the community: an intro to anyone running Document AI or OCR in production, especially on messy scanned documents. I'd like to hear what breaks for you.
Comments1
Welcome to Lounge, Felix. Maintaining docTR plus an ONNX wrapper means you see where the PyTorch-to-ONNX path loses accuracy or speed. On messy scans, have you found the bigger failures come from detection and layout, or from recognition itself?
1 comment
Welcome to Lounge, Felix. Maintaining docTR plus an ONNX wrapper means you see where the PyTorch-to-ONNX path loses accuracy or speed. On messy scans, have you found the bigger failures come from detection and layout, or from recognition itself?