Hire Utkarsh S. - AI Researcher

* Supporting a research project that applies advanced machine learning techniques to public health and legal policy analysis and active research on Influence Function. Key Contributions: * Spearheaded the development of a legal AI system using retrieval-augmented generation (RAG) for interpreting municipal codes across U.S. jurisdictions. * Designed a semantic search pipeline integrating OpenAI embeddings, PostgreSQL, and hybrid retrieval strategies (keyword + vector) for scalable legal query resolution. * Built a custom parser and chunking pipeline to convert large, unstructured statutes into structured, searchable formats for LLM inference. * Evaluated model performance using RAGAS, enabling quantitative and qualitative benchmarking of LLM-based legal summarization. * Conducted research on Deep Metric Learning (DML) to explore embedding space interpretability and its integration into legal document similarity search. * Investigated influence functions to trace model predictions back to influential training examples, supporting interpretability and debiasing efforts in legal NLP. * Active member of the Targeted Learning Study Group, exploring advanced statistical frameworks for causal inference using semi-parametric models and machine learning.

Backed by $6.5M seed funding at a $30M valuation, FileRead is building enterprise AI software to streamline legal research workflows. - Developed and deployed advanced NLP and LLM-based pipelines for classification and semantic analysis of complex legal and enterprise documents, achieving 95\% + F1 score and significantly improving retrieval efficiency for multimodal document intelligence systems. - Designed and optimized a production-grade Retrieval-Augmented Generation (RAG) pipeline, enabling FileRead’s flagship multimodal AI suite to deliver faster, more contextually accurate responses to legal research queries.

Selected as one of 13 graduate students from the U.S. and Canada for the highly competitive Japan-US Advanced Collaborative Education Program (JUACEP). Conducted research under Professor Toshiaki Fujii on Region of Interest (ROI)-based medical image compression for DICOM imaging. Designed a prototype pipeline to compress tumor regions less aggressively than surrounding areas, combining techniques from machine learning and medical imaging. Gained experience in cross-cultural research collaboration and presented findings to a multidisciplinary academic audience.