Yashwanth D

Built a role-based AI document assistant that dynamically routes user queries to a DuckDB-backed SQL engine or a Chroma-based RAG pipeline, enforcing RBAC policies and reducing unauthorized data exposure risk by 100%. Designed and implemented an LLM-driven NL→SQL agent with a sandboxed DuckDB execution layer and intelligent fallback to RAG, reducing structured-query failures and improving answer coverage. Implemented role-aware retrieval and integrated a Cohere reranker to optimize document ranking, improving top-3 retrieval precision by approximately 30%. Built an automated RAG evaluation pipeline that generates synthetic QA pairs and computes faithfulness, relevance, and conciseness using LLM-based scoring for continuous model validation. Delivered an end-to-end system with a FastAPI backend, Chroma embeddings, DuckDB analytics layer with CI-driven automated API endpoint testing reducing regression issues by approximately 40%. Designed and implemented Dockerized CI/CD pipelines using Jenkins and GitHub Actions to automate ML model validation and test suite execution on NPU platforms, enabling dynamic scheduling across environments and reducing manual testing effort by 60%.

Developed an automated insurance processing system leveraging advanced NLP techniques and OpenAI’s GPT-4 to optimize policy validation and accurately evaluate claim coverage. Processed large volumes of policy and claim data through robust preprocessing pipelines, utilizing spaCy for Named Entity Recognition (NER) and regex-based rules for data validation. Fine-tuned GPT-3.5 Turbo to enhance structured claim classification and decision-making efficiency, reducing inference costs while reserving GPT-4 for complex policy reasoning and edge-case evaluation. Successfully completed onboarding training at upGrad, excelling in SQL, Python, Machine Learning, Deep Learning, NLP, AWS achieving a score of 92%.