Email  /  GitHub  /  Google Scholar  /  LinkedIn

RL Post-Training for Efficient Edge AI

I am interested in post-training methods that make AI systems more efficient after the base model already exists. In EMBER and StoreAgent, this means moving beyond static retrieval and training memory policies that decide what evidence to retain, how to structure it, and how to recall it later.

This direction connects model behavior to systems constraints. Edge AI often has small memory budgets, expensive context windows, and limited compute. A useful model should learn policies that respect those constraints instead of assuming unbounded retrieval or full-history access.