I’m an AI researcher and engineer working on language models, AI agents, and computational creativity. I study how models generate, judge, and interact, combining model training and adaptation with rigorous evaluation. I build across the stack: from preparing data and designing experiments to developing inference pipelines and interactive applications people can use.
At IIIT Hyderabad’s Language Technologies Research Center (LTRC), I work with Prof. Radhika Mamidi on computational creativity and language model evaluation. My broader engineering work includes Mixture-of-Experts architectures, multi-agent learning systems, offline RAG, and generative storytelling.
Previously, as AI Research Lead at Telepathic Software, I led evaluation and analysis for The Humor Tax, studying how post-training affects humor quality and diversity through human evaluation and calibrated model judging.
How much does a model’s judgment change when the same text is presented differently? An audit of Jev, Clef Flash, and Perplexity Decider across humor, helpfulness, and summarization: 9,000 items, six datasets, and 702,000 recorded judgments. The release separates repeatability, presentation robustness, and agreement with people, and includes predictions and offline result reproduction.
Evaluating Humor Generation Across the Qwen3 Model Family. A study of how model scale relates to humor quality across five Qwen3 models, from 8B to 235B parameters, using human and automated evaluation.
GTVH-Guided LLM Decomposition · j10official at SemEval-2026 Task 1. A five-stage DSPy pipeline that translates humor theory into a structured generation process. Second place in both Chinese humor generation and GIF fill-in-the-blank.
Hugging Face Transformers · DSPy · PEFT / LoRA · RAG · Mixture-of-Experts
Research collaborations & applied AI projects
jatin.agrawal@research.iiit.ac.in
