Ucchwas Talukder Utsha

Advancing real-time medical care with embedded IoT, Cyber-Physical Systems, and edge-AI-driven machine learning.

I’m revolutionizing real-time health monitoring by integrating ultra-low-power wearable sensors with edge-AI and advanced machine learning, including Large Language Models (LLMs), to deliver proactive, personalized healthcare solutions. My research bridges cyber-physical systems, embedded IoT, and generative AI, creating scalable, secure, and impactful technologies that empower clinical decision-making. Here’s what drives my work:

  1. Real-Time Physiological Insights
  • Innovation: I’ve developed on-device algorithms achieving over 96.5% accuracy in detecting cardiac and respiratory anomalies, such as arrhythmias or irregular breathing patterns, using ECG and respiratory signals.
  • LLM Integration: Leveraging LLMs like LLaMA, I’ve built context-aware models that interpret physiological signals without extensive pre-training, enabling zero-shot anomaly detection and natural language-driven health insights. For example, my silent speech recognition system uses LLMs to decode unvoiced EMG signals, transforming them into actionable health data.
  • Hardware Edge: My inkjet-printed, flexible electrodes are designed for comfort and reusability, capturing high-fidelity signals and providing reliable data for continuous monitoring.
  1. Adversarial-Robust AI for Healthcare
    • Focus: I engineer machine learning models resilient to adversarial attacks, ensuring reliable health analytics in high-stakes environments.
    • LLM-Powered Defense: By integrating LLMs with adversarial training, I’ve developed systems that detect and mitigate boundary attacks in real time, enhancing model robustness for ECG classification. These models use LLM embeddings to contextualize low-confidence inputs, improving detection sensitivity by 7% for rare cardiac events.
  2. Scalable Wearable IoT Ecosystems
    • End-to-End Design: I’ve architected full-stack IoT systems, from custom sensor hardware (e.g., inkjet-printed electrodes, AD8232 for ECG data, IMU for respiratory data) connected via Bluetooth to a mobile app, to secure cloud integration for smooth patient monitoring.
    • Edge-AI Optimization: My pipelines deploy lightweight CNNs, LSTMs, and TensorFlow Lite models on resource-constrained devices, cutting inference latency by 40% while preserving data privacy through on-device processing.
    • GAN-Enhanced Monitoring: I’m exploring LLMs to generate synthetic physiological data via generative adversarial networks (GANs), addressing data scarcity for rare conditions. This approach, in conjunction with federated learning, ensures privacy-preserving and scalable health analytics, which is critical for cloud-based healthcare applications.

My research fuses AI innovation with practical engineering, harnessing LLMs to unlock new possibilities in healthcare. From generating synthetic ECG data to enabling conversational health assistants, LLMs amplify my systems’ ability to deliver personalized, real-time insights. I’m passionate about transforming complex algorithms into accessible tools that save lives, whether through wearables, smart diagnostics, or privacy-first platforms. My work is built to provide intelligent, secure, and user-centric technologies.