research

Research and engineering across foundation models, healthcare AI, and power systems.

AI for healthcare

Early detection of cognitive decline (Alzheimer’s disease, Mild Cognitive Impairment) from wearable and multimodal signals, at Connected Future Labs.

  • Automated pipelines that ingest, clean, and harmonize streaming PPG, EDA, accelerometer, skin temperature, Muse EEG, and IMU data.
  • Physiological biomarkers (HRV, sympathetic/parasympathetic responses, neural features) that separate MCI from healthy cohorts.
  • AWS S3 and SageMaker pipelines that audit data quality and detect signal anomalies for downstream classification.
  • Clinical decision-support with large language models; chapters on AI in diagnosis, treatment, and rheumatoid arthritis.

Foundation models and distributed machine learning

Federated learning, split learning, and decentralized multi-agent learning for heterogeneous edge devices, with convergence analysis in convex and non-convex settings.

  • Dynamic tiering for federated learning: 80% shorter training time on heterogeneous clients (IEEE Internet of Things Journal, 2024).
  • Communication-efficient workload balancing for decentralized learning: 26% less slower-side compute (IEEE ICDCS, 2024; 21.9% acceptance rate).
  • Privacy-preserving SplitFed with distance correlation and patch shuffling.
  • SOLAR, a subspace-based parameter-efficient fine-tuning method: 35% lower communication and storage than LoRA on ViT, LLaMA, and GPT-2 (under review).

AI for power systems

Federated and privacy-preserving machine learning for the grid and EV ecosystems, informed by nearly a decade of prior control and power/data center engineering experience.

  • Federated XGBoost and MLP classifiers for current, voltage, and frequency faults: 91% accuracy using 80% of local data.
  • Privacy-preserving federated event classification with additive homomorphic encryption: 3,877 events, 23 PMUs, 88–90% accuracy on highly non-IID data.
  • Feature collusion attacks on PMU data-driven event classifiers.
  • Generative AI and distributed ML for EV ecosystems (smart charging, V2G): +44.5% battery SOH estimation accuracy and 100% CAN bus anomaly detection in the surveyed frameworks.
  • Data center infrastructure engineering (2014–2022): redesigned power/cooling for a 22% reduction in energy consumption and led three builds worth $35M.