Artificial Intelligence
Computer Vision, multi-sensory threshold engines, and frame parsing mechanics.
AI-Powered Real-Time Gym Trainer Architecture
Architected and engineered an intelligent, computer vision-driven health and fitness web application deployed on Streamlit, delivering automated, real-time exercise form correction. Developed a high-performance video processing pipeline that seamlessly ingests live streams from local integrated webcams. Integrated MediaPipe pose estimation models with Python and OpenCV to map human skeletal frameworks, applying custom trigonometric algorithms to evaluate joint angles against biomechanical standards for instant user feedback.
Metrics at a Glance:
Key Architectural Achievements:
- Streamlit Architecture: Utilized the Streamlit framework to deploy a highly responsive, unified Python web interface, eliminating the need for complex decoupled API routing layers.
- Webcam Stream Ingestion: Built a dedicated local webcam stream handler utilizing OpenCV to capture and process live video inputs directly from the user's local hardware interface.
- Real-Time Form Correction: Engineered custom spatial geometry algorithms to calculate precise joint angles dynamically, providing instantaneous audio-visual feedback on exercise posture.
- Optimized Stream Processing: Structured an asynchronous video ingestion pipeline using OpenCV, maintaining a consistent 30+ FPS capture rate while minimizing CPU overhead on local devices.
- User Interface Design: Developed a simple, clutter-free web user interface featuring real-time telemetry rendering and layout panels for direct user interaction during workouts.
Artistic Style Transfer (AdaIN) Architecture
Architected and engineered a high-performance generative computer vision application that performs real-time arbitrary neural style transfer using Adaptive Instance Normalization (AdaIN). Implemented a decoupled encoder-decoder model structure in PyTorch, leveraging a normalized VGG-19 convolutional backbone to isolate structural geometry from artistic textures. Trained a deep reconstruction decoder from scratch using dynamic image transformations, delivering seamless blending of target brushstrokes, color palettes, and textures while strictly preserving original content layouts.
Metrics at a Glance:
Key Architectural Achievements:
- AdaIN Feature Alignment: Integrated an Adaptive Instance Normalization layer to dynamically align the mean and variance of content feature maps with target style statistics, enabling arbitrary style adaptation without retraining.
- Large-Scale Decoder Training: Scaled a PyTorch optimization pipeline over 200 epochs and 162,500 total steps, managing a combinatorial dataset of 13,000+ COCO content images and 2,000+ Kaggle Painter-by-Numbers style images.
- Multi-Objective Loss Tuning: Engineered a highly skewed 10:1 style-to-content loss function, coupled with an inverse-time learning rate decay scheduler to ensure stable structural convergence.
- High-Throughput Batch Processing: Optimized data ingestion using torchvision and Pillow to resize, randomly crop, and process a high-density throughput of 16 content-style image pairs concurrently per training iteration.
- Interactive UI Deployment: Wrapped the end-to-end model inference pipeline into an intuitive Gradio web dashboard, facilitating real-time user testing, style blending controls, and instant visual artifact evaluations.