Jaydev Tonde

Working on development of products feedback classification problem. The goal is to classify the given human feedback in 3 level of them. Trying with core transfom model. It is giving quite good results but the clients requirements is to gnerate the model explanations also. Possibly move to LLM inference.

Developed an AI Agents web crawling tool designed to parse clients' websites for specific requirements. This tool leverages Helium browser automation functions as tools for AI agent. The AI agents dynamically decide which subsequent links to crawl during the process and which data to save for user. I used huggingfaces mole agents for this.

The goal was to understand and retrieve citations relevant to the user's query, which is essentially a retrieval task. Next, key requirements were extracted from the relevant citations using GPT. After extracting the key requirements, a manual team verifies these requirements, followed by the generation of a summary of the citations. This summary is now being served to clients.

Tried fine-tuning the Qwen-2 7B VL model for extracting form fields related to borrower analytics. Initially, we attempted inference using prompts alone but did not achieve the expected performance, so we decided to proceed with fine-tuning. Our team currently has access to two H100 GPUs, which are being used for model inference.

The goal was to train a transformer model to classify the given collateral text into 36 classes/asset tags. We experimented with different types of models, ranging from XGBoost and Decision Trees to transformer encoder models like ModernBERT. Ultimately, we focused on fine-tuning large language models (LLMs) for multi-label classification. Models such as GPT-2, Phi, and Llama3.2 (1B) demonstrated

Developed an AI Agents utility to parse logs from the 4 Extraction APIs and identify issues that occurred in production for specific document uploads. The solution utilizes three AI agents: one for database access and another for analyzing Splunk API logs and onc agent as coordinator.

Developed a complete pipeline for extraction and classification using LLMs, which includes document classification, form taxonomy generation, form field extraction, and field validation. For this task, I used GPT-4o. The team now has a comprehensive prompt management system for each of the tasks mentioned above.

Worked on research and development of a domain-specific language parser using ANTLR and Python. This parser was completely developed in ANTLR-specific language and achieved approximately 95% accuracy. After development of this parser, numerous use cases emerged, and now this parser serves as the backbone for 2 products in the company.

Worked on model training for document classification using both NLP and image processing techniques. I experimented with different models from HuggingFace, including DeBERTa, RoBERTa, and LongFormer for NLP tasks, and Donut and ViT for image classification.

Trained a Doc2Vec model for document embedding generation and used logistic regression as the classification model. The task was a binary document classification problem, where the classes were 'Retail Installment Contract' vs. 'Others'.

Worked in Java and SQL for developing the automation scripts to develop test suites for web apps.

Grade: 8.6

Completed a Master's degree focusing on advanced topics in computer science. Achieved a CGPA of 8.6, demonstrating strong academic performance in the field.

Grade: 73.37%

Obtained a Bachelor's degree in Computer Science with a strong academic record. Achieved a percentage of 73.37%, indicating a solid foundation in computer science principles.