Jyotirmay Khavasi

Optimized production-deployed Vision Encoder-Decoder and Layout Transformer models through batch classification and data augmentation techniques resulting in a 4% improvement across all metrics. Converted models to ONNX format for a 40% reduction in evaluation time for clients. Involved in Fine-Tuning Layout Transformer models for classification and entity recognition on custom data, enabling accurate multi-class classification and precise extraction of information across diverse documents. Developed an OpenCV based Table Detection pipeline utilizing object detection, image segmentation, morphological operations, contour extraction and coordinate geometry to identify, segment and reconstruct tables from PDFs, enabling accurate extraction of tabular data. Building domain specific LLM agents using Advanced RAG techniques with Qdrant VectorDB through Llama-Index framework to provide accurate and contextually relevant responses on custom data.

Created a template for Reinforcement Learning using Advantage Actor Critic algorithm, configuring code to efficiently utilize parallel processing for spawning multiple environments and handling various Reinforcement Learning tasks, using interchangeable and flexible components of TorchRL. Researched methods to efficiently isolate and segment relevant video sections frame by frame which enabled to extract high-quality input data for the RL model. Enhanced CI/CD pipelines with GitHub workflows and enabled Docker containerization. Improved the configuration by centralizing YAML attributes, allowing command-line overrides, and helped in integrating with Google Fire and Hydra. Implemented code refactoring with Vue.js tags, enabling template inheritance and resulting in more than 1000-line reduction across templates.

Worked on a NLP Model which extracts Rules from given Text. Software converts and Processes Given Text to output the Mathematical Expressions of the Rules. Worked on custom BERT architecture which is used for Named Entity Recognition with a Recall of 88%. Employed Random Forests for Classification of a given sentence into ’Rule’ or ’not Rule’ with 92% Accuracy.

Relevant Coursework: Discrete Mathematics, Probability and Statistics, Data Structures, Artificial Intelligence, Operating Systems, Database Management, Software Engineering, Cloud Computing.