Jyotirmay Khavasi

• Experimenting with Fine-Tuning Transformer-based, open-vocabulary object detectors like Grounding DINO for text-prompted product localization, and combining it with Segment Anything (SAM) models to generate product masks for downstream validation (eg: flagging non-compliant text regions outside the product area). • Optimizing GenAI models by benchmarking alternative models, reasoning/evaluation strategies, running batch evaluations at scale, and improving efficiency via cost/token and prompt optimization.

Built a production RAG system (92% precision, 89% recall) using Azure Search Index and LangChain. Developed an agentic PowerBI/DAX analytics platform enabling business users to perform complex analyses through natural language (NL-to-DAX). Created OCR-based document intelligence pipelines with 85% extraction accuracy. Engineered a Databricks/PySpark pipeline harmonizing 15+ years of automotive data, automating forecasting workflows and reduced data preparation time by 80% and maintained 100% data integrity.

Optimized production Vision Encoder-Decoder and Layout Transformer models using batch classification and data augmentation, improving performance metrics by 4% and reducing inference time by 40% through ONNX conversion. Fine-tuned Layout Transformers for document classification and entity extraction. Developed an OpenCV-based table detection pipeline for PDF data extraction. Built domain-specific LLM agents using Advanced RAG, Qdrant VectorDB, and LlamaIndex for accurate contextual question answering.

Contributed to PyTorch-Ignite as a Google Summer of Code developer, building a scalable Advantage Actor-Critic (A2C) Reinforcement Learning template with parallelized environment execution using TorchRL. Researched video segmentation techniques for RL data generation, enhanced CI/CD through GitHub Actions and Docker, improved configuration management with Hydra and Google Fire, and refactored templates reducing codebase size by 1000+ lines.