Research

My primary research focus is agentic scientific discovery and autonomous laboratories: machine-learning systems that plan and run experiments, recover from failures, and use results to decide what to do next. I draw on AI agents, reinforcement learning, generative and foundation models, and ML systems, with computational biology as a central application area.

Autonomous Laboratories and Agentic Scientific Discovery

Autonomous laboratories need more than a scripted protocol. Instruments disagree on interfaces, runs fail in physical ways, and the next useful experiment often depends on the last curve or plate. I work on agentic systems that orchestrate multi-instrument workflows, monitor readiness and safety conditions, and decide whether to accept a result or intervene — for example compressing a concentration range after a saturated dose-response curve.

During the limited Model Hardware Standard (MHS) research preview (announced by Anthropic), I led Carnegie Mellon’s work using MHS so an AI agent could operate a liquid handler, plate reader, robotic arm, and cameras for serial-dilution dose-response experiments. The preview specification is provisional and model-agnostic; we used Claude as one agent. I built the drivers and orchestration layer. Read the post. Nature reported on the experiments on 24 September 2026.

The same questions show up in programmable cloud laboratories, where evaluating agents on real hardware is slow and expensive. I am building simPCL, a simulator and benchmark generator for studying planning, fault recovery, and closed-loop scientific discovery under controlled scenarios. Related work includes agentic fault recovery for orchestration of programmable cloud laboratories (work in progress) and a contributed talk at the AI Scientist Summer Workshop (Microsoft Research, 2026). I am interested in extending this line toward 3D laboratory simulation, sim-to-real transfer, and reinforcement learning in MuJoCo, including world models for those environments.

Generative and Foundation Models

I design generative models for sequences with controllable properties, including GANs for 5’ UTR optimization and transformer-based seq2seq models for protein-conditional RNA design. This work emphasizes reproducible baselines and validation against held-out predictive models.

Reinforcement Learning and Decision-Making

I apply reinforcement learning to decision problems such as selecting informative gene panels for senescence biomarker discovery. This involves policy optimization, structure-aware rewards, and evaluation under positive-unlabeled settings common in biological data.

ML Systems

I also work on LLM systems topics in coursework, including extending FlashAttention and PagedAttention to MiniTorch. See the projects page for details.