Technical Projects

Overview of software engineering and machine learning projects

Projects

Rammer

Video generation system that takes Reddit posts and comments and automatically converts them into narrated videos with visual elements (similar to those popular Reddit story videos on YouTube/TikTok). The system uses:

  • RabbitMQ for message queuing
  • Google Cloud Text-to-Speech for narration
  • Remotion for video rendering

Capable of creating both long-form videos and shorter clips suitable for platforms like TikTok and Instagram.

E-Commerce Chat Assistant

AI-powered live chat system that integrates directly with e-commerce stores to provide real-time shopping assistance and automated customer service. The system uses:

  • tRPC for type-safe client-server communication
  • WebSocket for real-time messaging
  • Shopify API integration for e-commerce operations
  • Next.js for web applications
  • T3 Stack (TypeScript, tRPC, Tailwind)
  • Drizzle for database operations

Built as four interconnected applications:

  • @web: Main web application and landing page
  • @chat: Customer-facing chat widget that lives on merchant sites
  • @shopify: Merchant admin dashboard (Shopify app)
  • @thunderbolt: Backend processing engine

The system aims to reduce friction in the online shopping experience by providing AI-powered shopping assistance directly within the merchant's website.

Face Detection with Transfer Learning

A PyTorch-based facial detection system that leverages transfer learning from a pre-trained ImageNet model to identify and localize faces in images. The system uses:

  • PyTorch for deep learning framework
  • Pre-trained ImageNet model as the backbone
  • WIDER FACE dataset for face detection training
  • Custom dataset handling for training
  • Matplotlib for visualization of results

Core components:

  • Model Architecture
    • Pre-trained ImageNet backbone for feature extraction
    • Custom regression head for bounding box prediction
  • Data Processing Pipeline
    • ImageNet-style normalization for transfer learning compatibility
    • Custom BBoxDataset class for training data management
  • Training Infrastructure
    • Fine-tuning pipeline for the pre-trained model
    • Bounding box regression loss

Self Study

  • Deep learning for coders by fast.ai Jeremy Howard
  • Andrew NG's deep learning specialization on coursera
  • Used transfer learning to turn a general purpose image classification model into a face detection model