Shubham Songire

Architected and deployed an end-to-end enterprise RAG system for 10,000+ complex clinical study documents using LangChain, OpenAI GPT4o, and OpenSearch, deployed on AWS, enabling contextual Q&A, document comparison, and reference-backed summarization to accelerate internal research workflows (evaluated using RAGAs). Reduced clinical documentation time by 60% by developing a Validation Plan Generator powered by LLMs (Azure OpenAI Claude 3.5), accelerating PSR (Periodic Safety Report) generation time from 11 weeks to 4 weeks. Developed a time series forecasting model to project Verzenio sales trends. Integrated model with a live dashboard used by leadership for strategic planning and market analysis. Built and deployed end-to-end machine learning models on Microsoft Fabric, leveraging Dataflows Gen2 and Lakehouse for centralized data ingestion and transformation, and Synapse Data Science Notebooks for feature engineering and model training using Python, pandas, scikit-learn, and PyTorch. Utilized MLflow integration for experiment tracking, model versioning, and registration, and deployed models as Fabric endpoints to serve REST APIs and integrate with Power BI dashboards for real-time analytics. Implemented automated retraining pipelines using Fabric Data Pipelines, ensuring continuous learning, and maintaining model accuracy in production. Won Lillys Best Project Award for our team project and received the Best Individual Achiever Award for Q1 2024.

Developed a computer vision solution using YOLO architecture with optimized thresholds to enhance Face Quality Assessment (FQA) accuracy in low-light conditions. Achieved a significant 8% improvement in FQA effectiveness under low-light environments, thereby enhancing the accuracy of the spoof detection algorithm's guardrails.

Developed an AI-driven border surveillance and autonomous target tracking system using YOLOv7, DeepSort, and LiDAR-based SLAM, enabling real-time human detection, tracking, and coordinate-based navigation in defense zones. Integrated camera and LiDAR modules on NVIDIA Orin board and synchronized tracking data with an autonomous navigation vehicle through ROS2 (NAV2) for real-time pursuit and control-room monitoring. Tech Stack: Python, YOLOv7, DeepSort, LiDAR, ROS2 (NAV2), GraspNet, NVIDIA Orin, OpenCV.

Participated and spread awareness about different activities conducted by IRC in our college.

Worked for 2 years in a company on different data science projects, developed multiple AI solutions across healthcare, education and research domains, including tumor detection from clinical images using CNNs, and sentiment analysis using BERT. Built graph interpretation models to extract data from charts (bar, pie, histogram) and generate natural language summaries using computer vision and NLP techniques. Implemented image-to-text generation pipelines combining ResNet for visual feature extraction and NLP models for sentence formation, delivering insights for academic and industry use cases.