I work on multi-agent reinforcement learning and physics-informed machine learning, with a focus on multi-agent dynamics, physical law, and calibrated uncertainty. Recent work spans adaptive opponent modeling, probabilistic hyperdimensional computing, and safety for autonomous UAV swarms.
University of Oxford research project on adversarial robustness of LLM/VLM safeguards, including activation-level monitors, VLM emergent-misalignment probes, and monitor-evasion stress tests.
Opponent-aware, model-based co-training for multi-agent reinforcement learning. The policy learns a model of other agents' behavior and plans against it as they adapt. Co-authored manuscript in preparation.
A probabilistic library that treats every hypervector as a posterior distribution. It adds closed-form moment propagation, calibrated predictions, and coverage-guaranteed prediction sets to hyperdimensional computing. Available on PyPI (pip install bayes-hdc) and archived on Zenodo.
A survey of safety, performance, and verification for AI-driven UAV swarms. The paper uses a two-axis assurance framework (lifecycle × failure source) to map dense areas and open gaps. NSF-funded.
Available on PyPI (pip install bayes-hdc) and archived on Zenodo.
A BCI that decodes error-related potentials (ErrP) from EEG and uses them as a reinforcement signal, letting an agent map neural activity to actions. EEG detection at ~76% within-subject; built with NeuroTech@USC, runner-up at the 2026 intercollegiate BCI Competition at UC Berkeley.
Real-time telemetry for USC's Formula SAE racecar. Live sensor data moves from the MoTeC data-acquisition hardware, through the telemetry radio, to a dashboard. The stack also includes role-scoped config and an LDX watcher that re-injects values MoTeC strips from its .ldx files.
Design and UI/UX for the Neurotech@USC club site: the visual system, responsive layout, and the member, project, and recruiting pages. Built with React, Vite, and Tailwind CSS.
Incoming summer research with Prof. Saikat Dutta (Cornell).
Robotic Embedded Systems Lab. Multi-agent reinforcement learning: adaptive opponent modeling, calibrated strategy belief, and belief-conditioned planning for adversarial co-training.
FastAPI backend and Next.js frontend serving 100+ restaurants across 20+ chains; built the automated testing and CI/CD pipeline.
AI Explainability & Accountability Lab. Interpretability and accountability methods for machine-learning systems. Also robustness testing for autonomous-driving perception.
Algorithmic safety for autonomous drone swarms with Prof. Yongxin Liu. This work led to the UAV-swarm safety survey above.
Led a 6-person team building a cable impedance tester; optimized quantum-cascade-laser controller SDKs.
A machine that is merely good enough can take the wheel, because we hand it over ourselves.
Systems thinking through the lens of what the model leaves out.
Why the dimensions our models drop are where black swans hide.