About Me
Hi, I’m Harsha, a PhD student at IOL.Learn, Zuse Institute Berlin. I work on efficient and reliable large-scale deep learning systems, especially LLMs and the systems built around them.
My research focuses on making training and inference less compute- and memory-intensive under real deployment constraints. I study how compression, adaptation, and inference-system design can make LLMs more practical to train, serve, and evaluate.
I also work on evaluation and trustworthy reporting for models modified for efficiency, especially when aggregate benchmark scores hide important changes in behavior.
Research interests: quantization, distillation, speculative decoding, disaggregated inference, distributed training, reproducible evaluation, multilingual LMs, local agentic systems.
News
- Three new preprints are out: I co-led Every Eval Ever, a schema and community repository for AI evaluation results; contributed to Evaluation Cards, an interpretive layer for AI evaluation reporting; and collaborated on What Do Evolutionary Coding Agents Evolve?, led by Nico Pelleriti.
- Tutorial on Every Eval Ever has been accepted @ FAccT 2026. The focus will be on the theme of Building Community-Governed AI Evaluation Infrastructure
- Organizing Shared Task at ACL on Every Eval Ever with the EvalEval Coalition from Feb 2026 - May 2026.
- Completed my MSc at Saarland University and started a PhD at Zuse Institute Berlin, where I will be advised by Prof. Sebastian Pokutta and Max Zimmer.
- Completed my internship at AWS successfully completing a project on disaggregated inference and quantization.
- Started as Applied Scientist Intern at the Scale org at AWS Tuebingen with Jonas Kuebler. I will be working on distributed LLM inference optimization using quantization.
- Practical on Federated Learning, presented at Deep Learning Indaba in Dakar, Senegal. Joint work with Andrej Jovanovic and Luca Powell.
- Will be starting as an Applied Scientist Intern at the AIRE Team at AWS Tuebingen in November. Started as a Research Assistant at D2, CVML at the Max Planck Institute for Informatics, advised by Dr. Jonas Fischer, working on Mechanistic Interpretability of fMRI + Vision models.
- Pre-print: Cyclic Sparse Training: is it enough? joint work led by Advait Gadhikar and advised by Dr. Rebekka Burkholz is now out.
- On The Fairness Impacts of Hardware Selection in Machine Learning - joint work in collaboration between Cohere For AI + RAISE Lab at the University of Virginia accepted as a poster at ICML 2024.
- Pre-print On The Fairness Impacts of Hardware Selection in Machine Learning - joint work in collaboration between Cohere For AI + RAISE Lab at the University of Virginia is out. Advised by Ferdinando Fioretto and Sara Hooker.
- Started as a Research Assistant at the Relational Machine Learning Lab, advised by Dr. Rebekka Burkholz, working on topics related to sparsity and lottery tickets.