AI Researcher & Systems Builder | Ph.D. Candidate in Reinforcement Learning & Neuroscience

I sit at the intersection of Theoretical Reinforcement Learning and Closed-Loop Systems Engineering. My work focuses on building resource-rational agents that can reason, reflect, and adapt in sparse-reward environments.

I am building metacognitive architectures—AI systems that don't just solve tasks but monitor their own uncertainty to allocate compute efficiently.

• Dual-Process RL: Modeling the arbitration between fast heuristics (System 1) and costly planning (System 2).
• Closed-Loop Pipelines: Architecting real-time fMRI neurofeedback systems that couple biological substrates with autonomous agents.
• Sparse Reward Learning: Developing "Insight" mechanisms that allow agents to discover structural shortcuts when external feedback is removed.

A modular RL framework for modeling metacognitive control.

• The Build: A custom Monte Carlo Tree Search (MCTS) engine integrated with a Model-Free heuristic agent.
• The Tech: Implemented a "Singleton Latch" mechanism for one-shot schema induction and a parallelized Bayesian optimization pipeline (scikit-optimize) for parameter fitting.
• The Result: Validated against human behavioral data ($N=49$), proving that "Insight" acts as a structural gatekeeper for resource-rational control.

End-to-end Deep RL agent for abstract reasoning.

• The Build: Architected a custom OpenAI Gym-compatible environment for the card game SET.
• The Tech: Trained a Deep Q-Network (DQN) to navigate high-dimensional state spaces, optimizing for optimal solution paths.

Real-time signal processing pipeline for causal neural modulation.

• The Build: Implemented a closed-loop system that decodes fMRI patterns in real-time (MVPA) to induce reward prediction errors in human subjects.
• The Impact: Demonstrated a causal link between induced neural states and improved problem-solving performance.

• Languages:
• ML & RL:
• Tools: