Pankaj Rathi

Fine-tuned vision-language models on robotic data to improve grasp detection and object recognition accuracy. Trained robot manipulation policies from teleoperation demos using imitation learning and behavioral cloning. Built automated labeling system with vision-language models to scale dataset generation with quality checks. Engineered grasp prediction features and tracked experiments to optimize model performance metrics.

Trained self-supervised models for ADAS perception using unlabeled driving data to reduce annotation costs. Built GDPR-compliant anonymization pipeline preserving visual features essential for model training. Deployed services with Docker and GitLab CI/CD on Kubernetes for automated model inference. Optimized existing preprocessing pipeline improving throughput by 30% through caching and parallelization.

Built OCR pipeline to extract structured property data from real estate PDFs, achieving 70% accuracy. Implemented SQL validation rules and incremental ETL workflows, ensuring consistent data warehouse quality. Developed Power BI dashboards with automated refresh, enabling real-time analytics for 15+ stakeholders. Deployed LangChain chatbot on Streamlit for natural language queries, reducing data access time by 40%.