Joined Phonely as an ML Researcher
Training and deploying voice agents for customers handling millions of calls a month.
Read moreML Researcher @ Phonely
Amit Kiran Rege
Hi! I am Amit Kiran Rege, an ML Researcher at Phonely, where I train and deploy voice agents for customers handling millions of calls a month. I graduated with a PhD in Computer Science from the University of Colorado Boulder, advised by Claire Monteleoni.
My research interest is in Reinforcement Learning and Human-AI collaboration broadly construed. My recent focus is applying reinforcement learning to LLM post-training, while maintaining interests in foundational RL theory and the foundations of interpretability.
I am also interested in startups and early-stage venture, and was previously an Associate at the Deming Center Venture Fund, where I helped source and evaluate early-stage Colorado startups.
Outside work, I enjoy soccer, cricket, music, and following current affairs.
Updates
Quick highlights for talks, releases, paper milestones, and travel.
Paper accepted at CDC 2026
Paper: "Learning from Local Walks on Dynamic Graphs with Bandit Feedback".
Read moreAccepted to Off Season II (Founders, Inc.)
Six-week summer builder program for student founders.
Read moreBest Paper Award Finalist at L4DC 2026
Paper: "Flickering Multi-Armed Bandits".
Read morePublications
Featured papers with links to PDFs, code, slides, and project pages.
Learning from Local Walks on Dynamic Graphs with Bandit Feedback
Accepted at the 2026 IEEE Conference on Decision and Control.
Multi-Agent Lipschitz Bandits
Accepted at the Twenty-Ninth International Conference on Artificial Intelligence and Statistics.
Flickering Multi-Armed Bandits
Accepted oral presentation at the 8th Annual Learning for Dynamics and Control Conference.
Blog Posts
Long-form technical notes and research explainers.
Under Construction
New writing is in progress
Coming soon!
Side Projects and Demos
Interactive prototypes and research engineering experiments.
HEARSAY
A voice agent's safety guardrail checks what the speech recognizer wrote down, not what the caller said. HEARSAY reads the guard code, finds the handful of sounds its promise depends on, and tests the hearing exactly there with generated phone audio whose truth is known. Play the clips: the caller says fifteen, the line says fifty, and the guard approves the wrong transfer. It ships the promise with an honest envelope, certified on these lines, broken on those, untested elsewhere.
Whisperfaster-whisperClaudePythonGitHub Pages
GLASSBOX
An AI agent is dropped into a game it has never seen and recovers the rules as runnable code, tested cell by cell against every recorded observation. Includes invented rule variants that exist in no training data. Every reconstruction is decompiled back to code and playable.
RAGScope
Static retrieval debugger that shows where RAG failed: chunk split, metadata filter, or reranker regression. Each case includes gold support spans, counterfactual fixes, and answer claim audits.
JavaScriptGitHub PagesAstroJSON
CV and Resume
Download my full academic CV or a one-page resume.