AI systems that adapt to real-world constraints.
We study how learning and computing systems can remain useful, private, secure, fair, robust, and resource-aware when data, devices, accelerators, networks, energy, water, and carbon are heterogeneous.
STAIL is organized around two research pillars connected by a common idea: adaptation under heterogeneity. We build methods that explicitly reason about constraints instead of assuming uniform data, devices, infrastructure, or environmental conditions.
Learning without centralizing raw data, while addressing statistical heterogeneity, corrupted or poisoned data, unreliable updates, unequal device capability, fairness, privacy, security, and participation cost.
Scheduling and measurement for AI and flexible computing loads under time- and location-varying water stress, carbon intensity, cost, accelerator availability, and service constraints.
Each project is part of the same larger agenda: making resource and trust constraints explicit in AI decision-making.
How should data centers reason about water when the consequence of one liter depends on where and when it is consumed?
We developed a stress-adjusted water (SAW) perspective that combines water consumption with spatial and temporal scarcity. The work separates on-site cooling from off-site electricity-generation impacts and models electricity-related water provenance.
Environmental scheduling only matters if the chosen action is actually executable on the AI infrastructure.
The controller reasons jointly about SAW, carbon, electricity cost, GPU type and memory, topology, model residency, checkpoint state, deadlines, inference routing, and latency SLOs.
What if environmentally inexpensive clients are not statistically representative?
SAW-Fed treats client participation as the controllable decision, combining environmental preference with participation fairness and importance-corrected aggregation.
Charging “overnight” is a clock-time rule. Environmental signals change hour by hour.
This funded student research project studies joint carbon- and stress-adjusted-water-aware EV charging using a full-year Texas/ERCOT replay and residential charging traces.
Heterogeneous clients should not be treated as identical workers.
FedASL adapts server-side aggregation to reduce the influence of unfavorable or corrupted client updates. FedSRC moves quality-aware control to clients. FairHetero addresses fairness when resource-constrained clients cannot execute the same model architecture.
Students can enter through reproducibility, datasets, visualization, or bounded algorithmic extensions, then grow toward owning experiments, systems components, research questions, and papers.
Assistant Professor of Computer Science, Texas Lutheran University.
Personal website →Carbon- and stress-adjusted-water-aware EV charging: experiments, data analysis, visualization, and manuscript development.
Carbon- and stress-adjusted-water-aware EV charging research.
2026Balancing Bits and Drops: Stress-Adjusted Water Management for Data Centers.ACM e-Energy
2025Empowering Clients: Self-Adaptive Federated Learning for Data Quality Challenges.IEEE EDGE
2024Hardware-Sensitive Fairness in Heterogeneous Federated LearningACM TOMPECS
2023Enabling Low-Cost Server-Level Power Monitoring in Data Centers Using Conducted EMI.ACM SenSys
2022Computationally Efficient Auto-Weighted Aggregation for Heterogeneous Federated Learning.IEEE EDGE · Code ↗
Students interested in AI/ML, systems, privacy, cybersecurity, sustainability, optimization, data analysis, or scientific visualization are welcome to reach out.
ztalukder@tlu.edu →