LLM inference engine with radio-related application.
An app that caps how many reels you send to and receive from a friend. Instead of sending reels directly, they go into a queue — your friend receives a limited number per day, and vice versa. Designed to prevent sensory overload from the endless stream of shared content.
Building a voxel engine from scratch in C++ and OpenGL that converts LiDAR point cloud data (LAS/PLY) into a navigable 3D environment. A robot runs 3D A* pathfinding through the voxel grid, with a planned RL agent (DQN/PPO) exported via ONNX as the final phase. Also integrating a 3D depth camera pipeline for real-world point cloud capture.
An AI agent accessible over WhatsApp.
AI-powered anomaly detection system using Multiple Instance Learning and Weakly Supervised Learning for real-time surveillance. Implemented real-time video feature extraction using Meta's pretrained 3D CNN and a Deep GRU-based network, with mixed precision training on local GPUs.
Knowledge graphs to explore drug-disease associations for repurposing from biomedical literature at IISER Bhopal, using PrimeKG dataset. Built ETL pipelines from scientific literature and integrated with Neo4J for efficient querying.
End-to-end Retrieval-Augmented Generation chatbot for NPTEL courses built at IIT Madras. Consolidated textbooks and YouTube lecture transcripts using a cross-source chunking strategy. Migrated from a prototype stack (Dash, Ollama, ChromaDB) to a production backend (FastAPI, vLLM, Milvus), reducing response latency from ~80s to ~13s under 100 concurrent requests.
Self-supervised learning framework using Vision Transformers (DINO) for efficient pathology image analysis and disease detection, reducing dependency on labeled medical data.