Projects

Glutton Storage

Developed a full-stack web application using ReactJS, NextJS, and Typescript, integrating with Appwrite, a backend-as-a-service (BaaS) platform, and implemented authentication and file management features
Implemented key features such as user authentication, file uploads, file management (view, rename, delete), file sharing, and a dynamic dashboard, enhancing user experience and productivity
Created a responsive dashboard with global search functionality displaying storage metrics and recent uploads while enabling file sharing and collaboration
Designed a responsive and modern UI using TailwindCSS and ShadCN, ensuring cross-device compatibility and optimal user engagement

End-to-end Summary of HTTP Live Streaming

Developed a comprehensive solution for extracting and summarizing content from HTTP Live Streaming (HLS) sources using Python and LLMs, enhancing the accessibility of streaming media
Engineered a pipeline converting m3u8 streams to mp3 format using youtube-dl, then leverages Gemini 1.5 flash for intelligent audio processing and key point extraction
Demonstrated ability to transform streaming data into concise summaries saving manual review time or content creators and researchers

Implemented SWIM-style P2P membership protocol reducing false positive detection by 50% with sub-second query latency
Developed Hybrid Distributed File System (HyDFS) with 3x replication and near-linear scalability
Engineered fault-tolerant stream processing framework with exactly-once delivery semantics

Aligned an open-source LLM with reasoning traces, enabling coherent decision-making and effective action in complex, dynamic environments
Generated reasoning traces using the ReAct and Reflexion technique from WebShop (simulated e-commerce website) to fine-tune the Llama-3.2 8B Instruct, achieving performance comparable to GPT-4o in navigation and responding in interactive settings

Developed security solution integrating LLM responses with fine-tuned DistilBERT model
Reduced unsafe responses by over 70% through custom curated dataset training