Ucchwas Talukder Utsha

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

A selection of my key projects:

Smart Health App

Tools: Java, Android Studio
Repo: Smart‑Health‑Application‑V4

  • Developed a mobile app to visualize real‑time ECG and respiratory signals, providing users with instant insights into vital signs while detecting cardiac irregularities and respiratory patterns.
  • Implemented algorithms to calculate heart and respiration rates and optimized the app for mobile devices to ensure smooth performance and reliable monitoring.

CardioHelp Visualizer

Tools: Node.js, JavaScript, MongoDB, GeoJSON
Repo: CardioHelp‑Visualizer

  • Visualizes the geographic distribution of cardiac patients using interactive maps, enabling users to explore real‑time patient data, regional trends, and demographic patterns in cardiovascular health.
  • Delivers personalized health guidance and educational resources, allowing users to gain actionable insights into their own risk factors while promoting greater awareness of heart health.

ECG Classification using Deep Learning

Tools: Python, Google Colab, TensorFlow
Repo: ECG‑Classification‑using‑ML‑models

  • Built ML models to classify different cardiac beats using the MIT‑BIH Arrhythmia database.
  • Introduced Generative Adversarial Networks (GANs) to generate synthetic samples for rare beat types, boosting model accuracy.

Silent Speech Recognition Using LLM

Tools: Python, PyTorch, LLAMA 3.2B
Repo: Silent‑Speech

  • Developed an EMG adapter module that transforms unvoiced EMG signals into embeddings interpretable by LLAMA 3.2B, enabling context‑aware, zero‑audio speech decoding on a closed‑vocabulary task.
  • Built robust preprocessing and fine‑tuning pipelines to seamlessly integrate EMG features with the LLM, demonstrating silent‑speech decoding with minimal training data.

AI‑Driven Defense Against Boundary Attacks

Tools: Python, PyTorch, Foolbox
Repo: AI‑Driven‑Defense‑Against‑Boundary‑Attacks

  • Combined adversarial training with boundary‐based detection to protect classifiers from black‑box boundary attacks while preserving high performance.
  • Designed automated Foolbox evaluation workflows to systematically stress‑test model resilience and trigger defenses on suspicious inputs.

On‑Demand Professor Q&A Bot

Tools: Python, OpenAI, Chainlit, Chroma DB
Repo: OnDemand‑Professor‑Q‑A‑Bot

  • Built a Q&A assistant leveraging vector‑based retrieval over lecture materials to help professors find slides and notes instantly.
  • Integrated GPT‑3.5 Turbo for real‑time interactions and used PDF loaders to index course materials.

Apnea‑Non‑Apnea Analysis

Tools: Python, TensorFlow, Google Colab
Repo: Apnea‑ECG

  • Developed ML pipelines to detect sleep apnea events from respiratory signals in the Apnea‑ECG database.
  • Extracted time‑ and frequency‑domain features to improve classification performance.

Sleep Stage Detection Using Polysomnographic Data

Tools: Python, Google Colab
Repo: Sleep-Stage-Detection

  • Analyzed combined ECG and respiration belt data from the MIT‑BIH Polysomnographic database.
  • Identified breathing patterns and cardiac events to classify sleep stages for sleep disorder research.

Adversarial Machine Learning

Tools: Python, PyTorch, Jupyter Notebook
Repo: Adversarial‑Machine‑Learning

  • Implemented untargeted and targeted attacks (FGSM, PGD, C&W) and defenses such as adversarial training.
  • Conducted experiments on image classification models to measure and improve robustness.

Compiler Design

Tools: C, C++, Flex, Bison
Repo: Compiler‑Offlines

  • Built lexical analyzers, parsers, and syntax checkers as part of multiple compiler‐design assignments.

Room Renting Platform

Tools: PHP, Laravel
Repo: Room_Renting

  • Created a web‑based booking platform where clients can search and reserve rooms, and owners can list properties and manage bookings.