Gautam Agrawal

About

Applied LLMs to classify long nonprofit mission statements into 8 segments without labeled data, using an ensemble of clustering/embedding methods plus zero-shot RoBERTa/BART and a Tree-of-Thought prompting pipeline with LLM-as-judge evaluation (Gemma). Also built LangChain/LlamaIndex agentic RAG workflows including a text-to-SQL data analysis assistant grounded on DB schema with retries and performance optimizations on an HPC cluster.

Key Strengths

• Built production-oriented LLM ensemble for zero-shot mission-statement classification without labeled data
• Combines classical NLP/ML (clustering/embeddings) with modern LLM techniques (ToT, LLM-as-judge) to improve accuracy
• Systematic prompt/rule evaluation (tested rules on ~100 samples; selected best-performing set)
• Reliability-focused agent design (schema grounding, few-shot for complex queries, retry-on-error feedback loop)
• Scales LLM workloads using HPC partitioning and batch processing (batches of ~1000) with token limits
• Effective stakeholder collaboration via iterative feedback loops during keyword engineering and validation

Experience

Education

Skills

Languages