Shashwat Pandey is a Machine Learning Engineer specializing in the design and development of reliable, scalable, and trustworthy large language model systems. His expertise spans generative AI, natural language processing, retrieval-augmented generation, uncertainty quantification, hallucination detection, MLOps, and production AI evaluation. He has experience owning machine learning systems across the complete lifecycle, including problem definition, model development, deployment, monitoring, evaluation, and observability.
Shashwat currently works as a Machine Learning Engineer at Zillow Group, where he leads AI-powered customer communication and call intelligence initiatives. His work has contributed to the AI-Generated Messages platform, which has supported approximately 165 partners, more than 100,000 messages, and approximately 9,000 weekly active users. He also led the development of automated follow-up email and SMS generation for Zillow Home Loans and helped transform call summarization capabilities into a productized machine learning platform.
At Zillow, Shashwat established evaluation frameworks covering more than 600 calls and 500 long-call summaries. His work improved summary quality by 28% and reduced the hallucination rate from 11.4% to 8.2%. He also developed production monitoring metrics, observability systems, and compliant event architectures to strengthen the reliability, governance, and safety of borrower-facing AI applications.
Previously, Shashwat worked as a Machine Learning Engineer at the University of California, where he architected and deployed Cruz Chat, the university’s first AI-powered chatbot. He developed its retrieval-augmented generation platform using Elasticsearch, FAISS, and LlamaIndex, processing more than 10,000 documents. He also deployed GPU-accelerated LLM inference systems on NVIDIA A100 infrastructure and improved retrieval latency, answer quality, experimentation speed, and conversational system performance.
Earlier in his career at Amdocs, Shashwat contributed to a large-scale Azure migration involving more than 50 billion subscriber records across over 30 enterprise systems. He developed churn prediction models that achieved 91% accuracy and supported a 48% reduction in churn. He also applied transformer-based models, including BERT and GPT-3, to customer sentiment analysis and optimized machine learning workloads using CUDA, reducing model training time by 50%.
Shashwat holds a Master of Science in Computer Science with a specialization in Natural Language Processing from the University of California, Santa Cruz, and a Bachelor of Science in Computer Engineering and Electronics from Jaypee Institute of Information Technology. His research focuses on the reliability of large language models—including uncertainty quantification for reasoning LLMs and independent, black-box auditing of deployed on-device models for calibration, confident confabulation, and over-refusal, together with mitigations that require no model access. His LLM research has been disseminated through arXiv and the IEEE International Conference on Computational Intelligence and Computing Applications.