Yash Mathur

Fine-tuned Llama-based LLMs using SFT and LoRA for legal document analysis using internal evaluation benchmarks. Integrated a reward model for ranking LLM outputs to improve consistency and reduce low-quality generations during RLHF training. Applied reinforcement learning methods (PPO, GRPO) on AWS SageMaker to improve model alignment and reduce hallucinations. Designed and deployed an agentic AI system automating legal document analysis workflows, reducing manual effort by ~30%. Built a RAG pipeline using embeddings and similarity-based retrieval over a 10k+ document corpus to improve retrieval relevance and answer quality. Managed ML infrastructure on AWS (SageMaker, EC2, S3, Lambda) and improved turnaround time for experimentation. Reduced document analysis time from ~8–10 minutes to ~3–4 minutes per document and used Weights & Biases for experiment tracking and reproducibility.

Built a RAG system using LlamaIndex and MongoDB. Developed AWS-based ingestion and preprocessing pipelines including embedding generation and storage using S3 and EC2. Designed embedding workflows for scalable document retrieval, improving retrieval accuracy by 15% through iterative tuning and evaluation. Optimised system performance for faster response times and scalability.

Developed a recommendation system using Random Forest, XGBoost, and LightGBM, improving F1 score by 11% and increasing user engagement by 8%. Built ML pipelines on AWS (SageMaker, Lambda, S3) for automated training and inference. Implemented CI/CD workflows using GitLab and processed large datasets using SQL and Snowflake.