Projects
Distributed Training Infrastructure
Automating AI training with fast, intelligent configuration
Containerized distributed system for rapid AI training and deployment. Built to get large-scale AI prototypes trained in minutes with easy cluster to edge computing transfer.
Hyperparameter Optimization Tool
Automating AI training with fast, intelligent configuration
Developed HyperController, a control-theoretic hyperparameter optimizer using Kalman filters and multi-armed bandits. Achieved 1000× speedup and state-of-the-art cumulative rewards in 4 out of 5 OpenAI Gym environments compared to leading baselines.
Bandit Portfolios for Dynamic Resource Allocation
Decision-making in dynamic financial systems
Designed algorithms for sequential decision-making in non-stationary environments using stochastic control and Bayesian inference. Applied Kalman filter-based bandit methods to optimize asset allocation and trading strategies under uncertainty. Additional work included anomaly detection and time-series forecasting for finance-adjacent systems.
IQT Lab41: Monitoring AI Systems in Production
Real-time anomaly detection for model alignment and safety
Developed a covariate shift detection algorithm to monitor real-time computer vision models (CIFAR-10), improving model reliability under distributional drift in deployed AI systems.
Computational Neuroscience at HHMI Janelia
Scaling simulations of neural circuits for systems neuroscience
Applied model reduction techniques to accelerate large-scale neural simulations of the fly brain’s visual system, enabling real-time experimentation in biologically inspired models.
Neural Models of Sleep at NYU
Modeling transitions in large-scale brain activity during sleep
Designed algorithms to optimize large neural populations in biologically inspired networks. Informed strategies for multi-agent coordination and decentralized control across structured brain states.