Projects
Redis From Scratch
Built a lightweight, in-memory key-value database from scratch in C, replicating core Redis functionality and underlying data structures. Engineered a concurrent TCP server using standard C socket programming to handle multiple client connections simultaneously. Parsed the REdis Serialization Protocol (RESP) to decode raw network byte streams into executable database commands. Developed support for essential commands (PING, ECHO, SET, GET) and implemented time-to-live (TTL) logic for automatic key expiration. Implemented master-replica synchronization, enabling replica nodes to connect, perform handshakes, and reliably synchronize state with the master server.
View ProjectVariational Autoencoder - QuickDraw Dataset
Trained a custom Variational Autoencoder (VAE) from scratch on the Google QuickDraw dataset to autonomously generate and interpolate hand-drawn sketches. Designed an advanced encoder-decoder architecture incorporating ResNet-style Residual Blocks (with SiLU activations) and Scaled Dot-Product Attention to capture long-range spatial dependencies. Implemented the reparameterization trick and a composite loss function, specifically utilizing KL Annealing to successfully prevent posterior collapse during training. Stabilized mixed-precision training by implementing gradient clipping to resolve gradient explosion issues. Engineered custom inference methods, utilizing Aggregate Posterior sampling rather than standard normal priors to significantly improve output quality and prevent blank generations.
View ProjectBhagavad Gita Conversational AI
Curated a custom, domain-specific dataset on the Bhagavad Gita from scratch by scraping websites, running OCR on scanned books, and transcribing YouTube audio. Leveraged the Gemini API to process raw, unstructured text and programmatically generate synthetic questionanswer pairs, yielding a high-quality dataset for LLM instruction tuning. Fine-tuned the Google Gemma-3-1B-IT model using Hugging Face libraries to create an AI capable of answering questions based on the text. Managed the training pipeline on rented GPUs from runpod.io and published the final model weights to the Hugging Face.
View ProjectShopmate
Engineered an AI-powered personal shopping assistant using a multi-step LangGraph agent to automate product discovery, autonomously comparing items and recommending the best options to reduce search time. Created an intent understanding module that analyzes user requests, conducts preliminary searches on product categories, and dynamically asks context-aware clarifying questions before initiating a deep search. Built scraping pipeline to extract and structure product details, pricing, and user reviews directly from Amazon, Flipkart, and Google Search. Designed a recommendation system that scores products based on user preferences and shows the top three matches for easy decisions.
View Project
Election Portal
Developed and maintained the official platform for conducting campus-wide elections at IIT Roorkee, enabling over 10K voters to elect Positions of Responsibility (PORs). Implemented robust security measures to ensure secure and tamper-proof voting processes. Wrote clean, extensible, and production-grade code across frontend and backend. Built using Next.js, Django, PostgreSQL, Strapi, and Docker. Managed scalable infrastructure with AWS and Terraform, ensuring high availability and fault tolerance.
View Project
IKS Console
Developed for the Indian Knowledge Systems (IKS) Division, Ministry of Education, Government of India Built a dynamic certificate generation platform that allows users to upload templates via a web console and generate personalized certificates from excel data. Automated the emailing of generated certificates directly to participants. Utilized Django, React.js, PostgreSQL, Docker, Pillow, and Nginx. Managed infrastructure including AWS SQS, Lambda, and S3 for scalable and efficient processing.
View Project