Designed custom GAN and Transformer-based Latent Diffusion architectures to generate high-fidelity synthetic ECG signals for medical data augmentation. Achieved 42% signal similarity improvement over classical GANs and outperformed U-Net LDMs by up to 95% across 8 evaluation metrics. Integrated the synthetic data into a downstream classification pipeline, improving heart disease detection accuracy by 20%.

Developed a real-time AI system combining emotion recognition, RAG-based LLM inference, and generative audio-visual media. Integrated memory-augmented feedback for dynamic emotional adaptation. Improved user emotional resonance by 56% and satisfaction by 67% in pilot studies, showcasing the power of personalized, adaptive multimedia environments.

Designed a custom game environment in Python using PyGame and implemented Deep RL and Adversarial RL agents to dynamically adapt game level difficulty based on player skill. Demonstrated the potential of reinforcement learning for immersive, user-driven game experience tuning.

Built an advanced image processing tool in Java using the MVC design pattern. Applied object-oriented principles to build a scalable and extensible framework for applying filters, transformations, and enhancements. Designed with future extensibility for UI components and additional I/O modules.

Designed a compact neural network architecture using MATLAB that accurately models both linear and non-linear logic gates. The architecture demonstrated robust performance on high-dimensional datasets and maintained high accuracy with significantly fewer neurons. Applicable to high-density, resource-constrained AI systems, this model improves training efficiency without sacrificing predictive capability.

Developed an efficient solver for the game Wordle using concepts from information theory and probability. The system selects guesses that maximize expected information gain, leveraging unigram frequency distributions to prioritize likely solutions. The method reliably solves Wordle puzzles in an average of 3–4 attempts, showcasing the power of entropy-based search in constrained decision spaces.

Focused on classifying urban built-up areas using neural networks trained on labeled ground truth datasets derived from K-means and Maximum Likelihood classifications. Integrated geographic data with neural architectures to automate land-use mapping across cities. The model demonstrated effective generalization across urban landscapes and provided scalable classification when applied on Google Earth Engine datasets.

Built a music recommendation system driven by real-time emotion detection using machine learning and OpenCV. The system captures facial expressions and classifies emotional states, recommending mood-aligned songs accordingly. Developed and trained the emotion detection model using deep learning techniques and integrated it with a dynamic music recommendation pipeline.

Developed an IoT-enabled wearable wristband capable of detecting user stress levels by analyzing biometric signals in real time. The system used neural networks to classify physiological data and provided feedback for workplace stress analysis. Integrated machine learning algorithms with embedded signal processing to deliver accurate and responsive monitoring in occupational settings.

Designed and implemented an embedded system to detect vehicular accidents and trigger real-time alerts using the AT89C51 microcontroller and Raspberry Pi. The system simulated crash detection in Proteus and utilized GPIO-based sensors to transmit location data upon impact. It served as a low-cost prototype for emergency response systems in smart vehicles.