Raghav Gupta

• Document Classification at Scale: Automated classification of medical documents using AWS Textract, Batch/Glue jobs, and SNS-SQS, reducing end-users' search time. System is able to handle large volumes efficiently.
• Appeal Prioritisation Model: Built an appeal scoring model with SHAP explainability, helping the reimbursement team prioritise appeals and maximise claim recovery. Both predictive and prescriptive in nature.
• Signature Extraction Project: Extracted patient signatures from TRF forms using segmentation models. Worked extensively to reduce error by 30%, important for manual cost cutting.
• Physician AI Agent: Built a general-purpose assistant for Physicians, using AWS Agents with a master router directing queries to sub-agents for patient data, reports and clinical trials. Integrated Lambdas as tools to Agent, added guardrails, custom prompts, and latency handling for secure and responsive user experiences. 
• Patient Report Knowledge Graph & Summarization: Modeled patient report data (JSON) into a clinical knowledge graph, linking gene alterations, therapies, and NCCN guidelines, enabling structured summaries for physicians.
• Sales AI Assistant: Designed a unified assistant to support sales reps by aggregating data across multiple systems—orders, cancellations, physician interactions, expenditures, and quota tracking. Enabled faster call prep, real-time insights, and a one-stop solution to improve sales planning and physician engagement.

Client: S3Global – Sports Analytics on Soccer Matches

• Built CV models for object detection, segmentation, pose-detection, and team classification on real-time camera feeds for European football clubs.
• Detr (ViT) Optimization: Achieved 30% speed enhancement using sparsity and 50% throughput improvement by applying mixed precision and TensorRT optimizations.
• HRNet for Pose Detection: Implemented pose visualization and correction using 3D triangulation and mmpose with COCO datasets.
• DeepStream Realtime Pipelines: Developed and optimized pipelines for real-time performance on various architectures, reducing latency significantly.
• Worked on event classification (e.g., pass, goal) and ported eventing pipelines to real-time.
• Expertise in camera syncing using OpenCV, Ffmpeg, Docker, and Redis for processing multiple camera feeds in sync.

With focus on Deep-Learning Models, worked at intersection of Computer Vision and Astrophysics.
Neural Ordinary Differential Equations (NODE), Variatonal Inference, Few-Shot Learning etc.

Publication: https://arxiv.org/abs/2012.07735 (33 Citations in 3 yrs)

Built Manager Dashboard from scratch. Millions of Orders fetched, displayed, sorted and filtered, all
real-time. Built using Flask, Postgres, SQLAlchemy and Tornado, with React front-end.
● pub/sub model: pushed data over multiple channels simultaneously (over web-sockets)
● filtering and sorting millions of orders using SQLAlchemy as ORM
● Prediction System: notify possible breach of order fulfilment

Deep Learning, Computer Vision, Machine Learning, Reinforcement Learning

3D Representations, Point-Clouds, MeshLab, Open3D, PIFu, NeRF

 BTech Computer Science + MSc Physics