Data Scientist, AXA Konzern AG
- Work on a production document-understanding pipeline that segments stacks of scanned pages into documents (boundary detection) and classifies each — hundreds of thousands of stacks per month — flagging low-confidence cases for human review.
- Built document-class-specific confidence thresholds that automate more documents while holding at least 95% precision per class, cutting manual review by around 30% in production.
- Researching pipeline upgrades: an LLM post-processing step for frequently-confused class pairs, and a single vision-language model (Qwen3.5-VL 4B) to replace the existing two-stage BERT+CLIP and RNN stack.
