Shubham Chauhan

Promoted from Data Science Intern. Launched an LLM-powered patient chatbot that collected symptoms, treatment history, and diagnostic context, cutting token usage by 70% through optimized rule-based routing. Engineered a RAG-based Help Section chatbot to automate responses for common user queries and intelligently route complex issues to customer service agents, deflecting 80% of support tickets. Trained and deployed a MaskRCNN segmentation model, improving acne and related skin condition detection accuracy by 6% in production. Developed a YOLOv11-based nodulocystic detection model using advanced computer vision techniques, achieving 0.43 mAP on clinical image datasets. Built an EfficientNet-based Melasma severity classifier, applying transfer learning and fine-tuning to achieve 86% accuracy on dermatological image data. Contributed to India's first segmentation-based baldness detection model, optimizing inference pipelines to achieve a 300 ms 95th-percentile response time. Designed a rule-based clinical decision-support model with medical experts, predicting treatment categories from user inputs and AI detections with 98% accuracy. Orchestrated auto-scalable deployments via FastAPI, Ray Serve, and TensorFlow Serving on Kubernetes, cutting cloud costs by 63% while sustaining sub-3-second response times. Established internal tooling for model performance monitoring, distribution shift detection, and post-deployment production tracking across AI systems. Created an AI-assisted annotation framework with medical experts, boosting annotation speed by 30% and improving annotator agreement for classification from 63% to 78%. Applied PyTorch and TensorFlow across model development workflows, including training, fine-tuning, and optimizing deep learning models for production deployment.

Built face-tracking and skin-analysis pipelines using SOTA computer vision models (MediaPipe, 3DDFA) for automated skin condition reporting, reducing manual review time by 40%. Designed lesion-overlap and severity-scoring metrics to aggregate model outputs across multiple images over defined timeframes. Developed tools to consolidate multi-annotator labels using IoU, Cohen-Kappa, Fleiss-Kappa, and other statistical agreement methods, improving label consistency by 23%.

Built an end-to-end Number Plate Recognition system using YOLOv5 and pytesseract, achieving 93% accuracy across diverse Indian plate formats. Evaluated camera types and configurations to minimize distortion and improve field performance of deployed vision models. Deployed a Flask-based web application for the NPR system, providing an accessible and user-friendly interface.