Papers, preprints, and
a couple of course projects.

Investigates geometric coherence properties of global model aggregation in federated Graph Neural Networks. Analyzes how heterogeneity in local graph structures impacts aggregation fidelity and proposes corrections to maintain geometric alignment across clients in distributed GNN training.

Addresses the challenge of modeling execution-level semantics in programming languages with structural constraints. Proposes a framework for capturing the semantic behavior of programs beyond syntax, to help with code analysis and generation tasks in NLP systems.

Presents DP-EMAR, a differentially private framework that detects and repairs corrupted or drifted model weights in federated IoT deployments. Uses calibrated noise mechanisms to preserve formal privacy guarantees while allowing aggregation to self-correct without exposing client data.

Explores quantum aggregation protocols on NISQ (Noisy Intermediate-Scale Quantum) devices for federated learning in Autonomous Driving Assistance Systems. Proposes noise-resilient quantum circuits and compares them against classical baselines under realistic hardware noise.

A course-project paper comparing human visual categorization strategies against ML models, looking at where and why they diverge. Sketches a neuro-symbolic framework combining connectionist perception with symbolic reasoning to help narrow that gap.

A team course project (four of us, Digital Health Equity at Georgia Tech) building a simulation that models how interventions like housing, nutrition, and employment support ripple through community health outcomes. Meant as a policy discussion tool, not a validated clinical model.

An undergraduate-thesis paper proposing an authentication method that maps deliberate gaze patterns to Morse code, read in real time by a simple AI tracker — aimed at hands-free authentication for users with motor impairments.

An undergraduate project on a FIFO buffer design for low-latency embedded video streaming — looking at throughput bottlenecks on constrained hardware and simple buffering strategies to keep the stream stable under variable network conditions.