About Me

Hello, and welcome to my website! I am Kuldeep Singh, a Senior Machine Learning Engineer with 7+ years of industry experience building and deploying NLP and LLM systems. I currently work at Quantiphi, on-site with client AT&T, where I focus on LLM post-training and evaluation. I fine-tune open-source models that replace GPT-4o and GPT-4.1 in production workflows, beating them by 5 to 10% on use-case metrics and saving roughly $800K/year. I also owned the integration of reinforcement learning (GRPO, PPO, GSPO) into AT&T’s Finetuning-as-a-Service platform and architected an evaluation microservice (LLM-as-a-judge and code-based judges) now used by 10+ teams. My research and engineering interests center on LLM post-training and reinforcement learning, evaluation methodology, Retrieval-Augmented Generation, and agentic systems.

Before Quantiphi, I was an AI Engineer at BYO, where I built RAG agents with guardrails and reproduced TRACE continual learning results on Llama-3.1-8B, and before that an AI Consultant with the Department of Family Medicine at Michigan State University, where I improved retrieval recall in a clinical RAG system by 22% using AWS Bedrock and OpenSearch. That MSU work grew out of my Master of Science in Data Science there (GPA 4.00/4.00), where I collaborated with Professors Ling Wang and Jiayu Zhou on LLM-based evaluation of clinical notes and Open Domain Question Answering, and completed a capstone under Professor Mohammad Ghassemi on predicting funding flows in public and private markets. I hold a Bachelor’s in Civil Engineering from Indian Institute of Technology Roorkee.

During summer 2023, I was an AI Research Associate Intern at JP Morgan & Chase, working under Simerjot Kaur and Charese Smiley on a financial question-answering system that reached SOTA results on FinQA. That work became my “FinQAPT” paper, which won Best Paper Runner-Up at ACM ICAIF 2024.

Earlier in my career I worked at ConcertAI, Bengaluru and Innoplexus Consulting Pvt. Ltd., Pune. At ConcertAI, I built information extraction models for Electronic Health Records of cancer patients, using weak supervision (SkWeak, Snorkel) to scale training data. At Innoplexus, I built NER and relation classification models on biomedical data and helped construct a biomedical knowledge base for drug discovery pipelines.

For the full breakdown of roles, publications, and skills, see my CV.