Mikołaj Dziewiatowski

At Data4Cure, Inc., I built an end-to-end quality-control application for biomedical NLP output and deployed it to production, where its results updated the live platform data.

I designed metrics for RAG-generated summaries, including checks for repetition, copied sentences, conciseness, hallucinations, and source annotation. I used them to benchmark open-source LLMs and select the model that regenerated production summaries, and built a pipeline for iterative prompt optimization.

For relation classification, I evaluated complementary prompts and models, then replaced the original pipeline with an open-source decision model. Production QC ran 6× faster at 99% precision, with a similar gain of about 10% in correct classifications.

At the University of Warsaw, I researched links between mutational signatures and histopathology for HistoSearch. I built a metadata-aware image similarity search using pathology foundation-model embeddings and FAISS, and released the pipeline and Gradio app on GitHub.