I design trustworthy and sustainable AI systems that operate under real constraints on privacy, security, fairness, hardware, communication, energy, water, and carbon.
My work asks one systems question at different scales: how should AI adapt when data, devices, accelerators, networks, geography, and environmental impacts are not uniform?
Trustworthiness and sustainability are not afterthoughts. I treat them as constraints that shape learning and systems decisions from the beginning.
Secure, privacy-preserving and robust learning under heterogeneous data, devices, reliability, and participation costs—including corrupted or poisoned client data and unreliable updates.
Resource-aware computing that jointly reasons about water stress, carbon, cost, accelerators, and operational feasibility.
Environmentally aware federated learning where client selection balances learning utility, participation fairness, environmental cost, and objective fidelity—without moving private data.
Stress-adjusted water accounting for data centers, separating on-site cooling from electricity-related water impacts and making geographic and temporal water stress part of operational decisions.
Hardware-Sensitive Fairness in Heterogeneous Federated Learning addresses fairness when resource-constrained clients cannot all execute the same model architecture.
Online scheduling for geo-distributed AI that considers SAW, carbon, electricity cost, GPU type and memory, topology, model residency, checkpoints, deadlines, and inference SLOs.
Joint carbon- and water-stress-aware EV charging using Texas/ERCOT signals and residential charging traces, extending sustainable scheduling beyond computing clusters.
Robust federated learning against unfavorable or corrupted client data: FedASL adapts server-side aggregation to reduce harmful updates, while FedSRC lets clients self-regulate training and upload based on local data quality.
Server-level power monitoring using conducted electromagnetic interference, connecting sustainable systems optimization to practical measurement.
Cybersecurity is one of the foundations of my trustworthy-AI work: understanding malicious, corrupted, unreliable, or privacy-sensitive participants and designing learning systems that remain useful under those conditions.
FedASL reduces the influence of unfavorable client updates under data corruption, while FedSRC moves quality-aware decisions to the client.
I want students to reason about a problem, explain why a method works, implement or evaluate it, and recognize the assumptions behind a technical choice.
01 Make abstract ideas visible and active.
02 Scaffold rigor without lowering expectations.
03 Evaluate computing, not merely use it.
04 Treat research mentorship as an extension of teaching.
STAIL connects secure and trustworthy AI, privacy-preserving and robust federated learning, sustainable AI systems, and undergraduate research.
Enter STAIL Lab →Texas Lutheran University · approximately $10,000
Supported undergraduate research, faculty effort, and summer housing for joint carbon- and stress-adjusted-water-aware EV charging.
External proposal direction spanning measurement, forecasting, water provenance, GPU/workload flexibility, carbon-aware operation, and AI-infrastructure decision-making.
Research and proposal contributions to NSF-supported work on sustainable data centers and conducted-EMI-based server-level power monitoring.
MLCommons Rising Stars · NVIDIA Headquarters · Santa Clara, California
Ph.D. Lightning Talk · The University of Texas at Arlington
Massachusetts General Hospital / Harvard Medical School