Atharva Ingle

Core contributor to Content Rocket, focusing on monetizing existing data and transforming user interactions with GenAI. Led multiple sub-projects and developed a standardized AI workflow for RAG, document classification, and chatbots, significantly reducing manual effort.

Built a search engine to recommend required licenses for starting a business, developing a HyDE workflow to expand user queries. Used synthetically generated descriptions and a reranker, achieving 96% retrieval accuracy.

Led AI capabilities for Borrower Analytics 2.0, processing 65 million Uniform Commercial Code (UCC) filings and fine-tuning a Qwen 2 VL 7B Vision-Language Model (VLM) for 95% extraction accuracy. Developed an LLM-as-a-judge framework to generate high-quality labeled data, significantly reducing human labeling effort for VLM fine-tuning.

Developed a sophisticated RAG-based chatbot leveraging an XML-based IRA knowledge base, achieving a 97% customer satisfaction rate. Designed and implemented an advanced RAG pipeline incorporating query transformation, parent-child retrieval, hierarchical indexing, and multi-vector database retrieval using LanceDB.

Built a RAG-based chatbot on 91k legal citations for efficient legal query resolution, enhancing retrieval with metadata filtering and query transformation. Used BERTTopic to generate hypothetical topics for improved search space reduction and query handling.

Developed an end-to-end internal framework for fine-tuning various production models like Donut, LiLT, and Qwen, making it accessible for non-technical users. Integrated MLflow for experiment tracking, model registry, and dataset/model versioning, supporting distributed multi-GPU training and advanced techniques.

Optimized Kaggle notebooks with added support for W&B tracking and monitoring tools. Wrote technical reports showcasing W&B features in diverse areas including medical imaging, Visual-Language Models, few-shot learning, PyTorch 2.0, HuggingFace, RAG, and LLMs.

Automated key information extraction (KIE) from legal documents, creating an in-house solution using transformer models like Donut, LiLT, and pix2struct. This achieved a 98% cost reduction compared to Azure Form Recognizer and included building a document classification pipeline.

Grade: 9.72/10

Completed a Bachelor of Technology in Instrumentation and Control Engineering. Studied core concepts and applied principles in various projects. Achieved a strong academic record with a CGPA of 9.72/10.