At Inspira, I build AI systems for a legal assistant used by 17k+ legal professionals. I shipped a citation-provenance gate that requires each legal reference to resolve to a retrieved source, after measuring unbacked references in production.
I extended the hybrid-search pipeline with vector and keyword retrieval, reranking, and query rewriting, and built document RAG with PDF parsing, Redis vector search, and page-level citations. I also developed an LLM-as-Judge pipeline to score production turns.
At Pixeon, I improved mammography lesion segmentation from 75% to 85% Dice using PyTorch, architecture refinements, and dataset curation. I also worked on pediatric bone-age estimation and production inference pipelines.
At Sensix Agro, I developed weed segmentation and classification over UAV orthomosaics and delivered shapefiles to target areas needing herbicide. My research includes published work on weed detection and high-voltage insulator imagery.
