Projects
Research projects and open-source contributions at the intersection of geometric deep learning and ML engineering
Featured Project
Featured
## Overview
Built an end-to-end machine learning pipeline for video generation using equivariant variational autoencoders (VAEs), implementing geometric deep learning principles to preserve symmetries in video data.
## Key Features
- **Modular ML Pipeline**: Implemented using PyTorch Lightning with configurable data loaders and architecture support
- **Equivariant Architecture**: Designed VAE layers that respect SO(3) rotational symmetries in video frames
- **Automated Monitoring**: Built-in experiment tracking and performance metrics using TensorBoard and MLflow
- **Production Deployment**: FastAPI-based inference service deployed on AWS infrastructure
- **Performance Optimization**: Achieved 7× inference speed improvement through model quantization and batching
## Technical Details
### Architecture
The equivariant VAE uses group-convolutional layers to maintain rotational equivariance throughout the encoding and decoding process. This allows the model to generalize better across different orientations of objects in video frames.
### Infrastructure
- **Training**: PyTorch Lightning on multi-GPU setup with distributed data parallel
- **Inference**: FastAPI microservice with async request handling
- **Deployment**: Docker containers orchestrated on AWS ECS
- **Monitoring**: Real-time metrics dashboard with Grafana
### Performance
- 7× faster inference compared to baseline implementation
- 40% reduction in model parameters through architectural improvements
- Maintained comparable reconstruction quality (PSNR: 28.5 dB)
## Impact
This project demonstrates the practical application of geometric deep learning principles in production ML systems, combining theoretical rigor with engineering best practices for scalable, maintainable code.
Equivariant VAE for Video Generation
End-to-end ML pipeline for video generation using equivariant variational autoencoders with modular architecture and automated monitoring
PyTorchPythonFastAPIDockerAWSVAEGeometric Deep Learning
## Overview
Built an end-to-end machine learning pipeline for video generation using equivariant variational autoencoders (VAEs), implementing geometric deep learning principles to preserve symmetries in video data.
## Key Features
- **Modular ML Pipeline**: Implemented using PyTorch Lightning with configurable data loaders and architecture support
- **Equivariant Architecture**: Designed VAE layers that respect SO(3) rotational symmetries in video frames
- **Automated Monitoring**: Built-in experiment tracking and performance metrics using TensorBoard and MLflow
- **Production Deployment**: FastAPI-based inference service deployed on AWS infrastructure
- **Performance Optimization**: Achieved 7× inference speed improvement through model quantization and batching
## Technical Details
### Architecture
The equivariant VAE uses group-convolutional layers to maintain rotational equivariance throughout the encoding and decoding process. This allows the model to generalize better across different orientations of objects in video frames.
### Infrastructure
- **Training**: PyTorch Lightning on multi-GPU setup with distributed data parallel
- **Inference**: FastAPI microservice with async request handling
- **Deployment**: Docker containers orchestrated on AWS ECS
- **Monitoring**: Real-time metrics dashboard with Grafana
### Performance
- 7× faster inference compared to baseline implementation
- 40% reduction in model parameters through architectural improvements
- Maintained comparable reconstruction quality (PSNR: 28.5 dB)
## Impact
This project demonstrates the practical application of geometric deep learning principles in production ML systems, combining theoretical rigor with engineering best practices for scalable, maintainable code.