☕️
I have over four years of experience working with computer vision and natural language processing. Currently, I work as a Machine Learning Engineer at ExB, where I am responsible for implementing algorithms for the task of unstructured document analysis, information extraction, and recognition.
Previously, as an academic member at Furtwangen University, I developed and researched on AI- and NLP-enhanced chatbots for (just-in-time) adaptive interventions.
Further, during a research fellowship, I created a spelling correction Benchmark, but my primary focus was creating and researching on NSEC, a neural network enhanced spelling error correction system combining statistical methods with deep learning methods. I not only used known Language Models like BERT, but also developed my own ones that better cope with the nuances of small typos and had a primary focus on XAI systems.
Skills
Experience
- Regression model for predicting the most probable table structure by representing everything as row- and col-spanning cells
- Using Graph Neural Networks for ICD-10 code classification in a hierarchical structured space of up to 18.000 labels
- Implemented a prototype of a hierarchical multi-label classification approach for detection structured information like address-fields and their internal structure (name, street, etc) in a single run.
- Introduced concepts of how to speed up machine learning models through quantization, model distillation and other features leading to smaller models and higher computation throughput.
- Build automatic dataset generation tool which simultaneously builds ground truth information for the tasks of document type detection, layout detection and segmentation, text recognition on paragraph-, word-, and character-level, including all required information like position and font information. Documents are created randomly through a templating process.
- Collected statistics for custom label NER task and made them available in human- and machine-readable formats.
Lecturer for the courses:
- Introduction to the Linux operation system
- Introduction to the Python programming language
- Gave “LaTeX Workshop” at the October Workshops, established by the Cognitive Science Departure
- Tutoring courses Algorithms and Datastructures, Pattern Recognition and Computer Graphics, Introduction to the Programming Language C++
- Started creating a new, modern, and modularized open dialog and intervention platform which can take advantage of modern NLP and NLU mechanisms in the background and allowing to produce open, out-of-place dialogues. The platform was separated in the backend which takes care about the dialogue management and the frontend which can be an Android app or the web browser itself.
- Maintained system using for the creation of interventions on an online platform
Research on the topics of Tokenization Repair and Spelling & Grammar Error Correction
- Iterative multi-stage approach on finding and fixing errors
- Memory and computation time optimizations on LSTMs
- Deep Language Model, based on LSTMs with stable results, even after > 20 stacked layers
- Developed a spelling error correction benchmark with different error categories and comparison on large set of available tools
- Started development on an automatic LaTeX-based corpus generation tool for the tasks of Layout Detection, Text Detection, and Text Extraction
Education
Thesis: Using half precision data representation in deep convolutional neural networks
Subsidiary subject: Cognitive Science
Thesis: Visual tracking of rodant feet
Subsidiary subject: Mathematics
Accomplishments
