Human feedback and preference learning
Designing more efficient ways to collect and use feedback when comparisons, rankings, and human labels are expensive.
AI research scientist and applied systems builder
Kousha Kalantari
I work on AI systems at the boundary between careful theory and useful products: learning from human feedback, data-efficient adaptation, recommender systems, and privacy-aware machine learning.
My path has moved through statistical learning, information theory, and large-scale product work. I care about AI that is technically grounded, practical in deployment, and shaped with humility about what these systems can and cannot do.
Current focus
Research that can travel from paper to product.
I am an Applied Scientist at AWS AI Labs and previously worked as a Research Scientist at Meta. Across academic and industry settings, I have worked on machine learning systems that connect statistical ideas with production constraints, team judgment, and real users.
Work
Where my work tends to live
Data-efficient model adaptation
Selecting informative examples for supervised fine-tuning and studying when smaller, better-chosen datasets can move models.
Recommender systems and bandits
Building and analyzing learning systems that make good decisions under uncertainty, limited feedback, and production pressure.
Privacy and information theory
Studying privacy-utility tradeoffs with the language of information, distortion, and statistical structure.
Selected publications
Recent and representative research
FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain
Comparing Few to Rank Many: Active Human Preference Learning Using Randomized Frank-Wolfe Method
Data-Efficient Supervised Fine-Tuning of Language Models Using Optimal Design
Optimal Design for Human Preference Elicitation
Optimal Design for Human Feedback
Fixed-Budget Best-Arm Identification with Heterogeneous Reward Variances
Robust Privacy-Utility Tradeoffs Under Differential Privacy and Hamming Distortion
Optimal Differential Privacy Mechanisms under Hamming Distortion
On the Fine Asymptotics of Information Theoretic Privacy
Background
A few coordinates
Applied Scientist
Current work on next-generation AI systems and applied machine learning research.
Research Scientist
Industry research experience across large-scale machine learning and product contexts.
Academic formation
PhD candidate at Caltech, currently on leave; MS in Electrical Engineering from Arizona State University; BS from Sharif University of Technology.
Contact