here are some of my projects — things i built, experimented with, or found interesting.
Smollocal is a full-stack multi-tenant web platform built to help local sellers manage their business and customers more efficiently. It provides two separate portals — one for sellers and one for customers — connected through a scalable backend and designed for real-time operations.
It’s optimized for modern browsers, mobile-ready, and crafted with a clean component-based architecture for easy extension and future integrations.
Key Features
Seller Portal
- Secure authentication with JWT
- Add/update/delete products with image uploads
- Manage orders with approval/denial functionality
- Add and view customers and their order history
- Category management for products
- Auto-generated QR codes for store links
- WhatsApp messaging integration
- Dashboard with real-time analytics
Customer Portal
- Access seller-specific storefronts via unique links or QR codes
- Browse products and place orders through a seamless checkout
- Track order status
- Dark mode toggle for better UX
This project implements SimCLR, a modern self-supervised learning framework, from scratch using PyTorch. The model learns meaningful visual representations without using labels by contrasting different augmented views of the same image.
The learned representations are evaluated using a linear probing protocol, demonstrating that the encoder captures non-trivial semantic structure purely through contrastive learning.
Key Components
Self-Supervised Training
- Custom SimCLR architecture with ResNet-18 encoder
- Projection head for contrastive embedding space
- Strong data augmentations for positive pair creation
- NT-Xent contrastive loss with temperature scaling
- Trained without using class labels
Representation Evaluation
- Frozen encoder with linear classifier on top
- Feature extraction from learned representations
- Linear probing on CIFAR-10
- Achieved ~25% accuracy with limited training
- Embedding visualization using t-SNE
Streaming Log Anomaly Detection is an AI-powered real-time monitoring system that classifies and explains log events using machine learning and large language models. It continuously listens to incoming logs, detects anomalies, and provides natural-language insights to help developers identify and resolve issues quickly.
Built with Pathway, Hugging Face Transformers, and Google Gemini, it’s designed for real-time processing, scalability, and easy integration into observability pipelines or DevOps dashboards.
Key Features
Log Processing Engine
- Streams logs in real-time using Pathway
- AI-powered classification into normal, warning, or critical
- Gemini integration for root-cause explanations
- Supports schema-flexible log formats
- Writes structured results to JSON outputs
- ETL-style pipeline: Extract, Transform, Load
Visualization Dashboard
- Built with Streamlit for real-time log visualization
- Filter and search by severity or timestamp
- View AI-generated explanations for anomalies
- Minimal, responsive, and dark-mode compatible UI
- Ready for deployment or demo on Streamlit Cloud