Projects
Text2Chair: AI-Powered 3D Furniture Generation in Mixed Reality
Developed end-to-end text-to-3D generative pipeline for Meta Quest 3, enabling real-time furniture visualization in mixed reality environments. Integrated Stable Diffusion XL Lightning with TripoSR image-to-mesh model via ComfyUI backend, achieving 15-20 second generation time from text prompt to interactive 3D object. Implemented custom OBJ loader with vertex color preservation, XR interaction system with physics-based manipulation, and wireless API communication via ngrok tunnel. Overcame technical challenges in URP shader vertex color rendering and MR collision detection, resulting in spatially-aware furniture placement with natural grab-and-throw mechanics.
CSXL: AI-Powered Interview Platform
Developed a production-ready RESTful service using FastAPI and Azure OpenAI to generate adaptive coding interview questions. Established a robust CI/CD pipeline with GitHub Actions and Docker to automate testing and deployment to OKD cloud, ensuring code quality for a collaborative 4-person team. Designed modular PostgreSQL schemas to support efficient content versioning.
Pikachu Dodgeball: MIPS-Based Game on FPGA
Designed and implemented a complete hardware-software system from scratch: built a single-cycle MIPS processor in Verilog, integrated memory-mapped I/O devices (VGA display controller, keyboard, accelerometer, sound, LEDs), and developed a physics-based dodgeball game entirely in MIPS Assembly. Implemented collision detection, gravity simulation, and sprite-based rendering on FPGA hardware, demonstrating end-to-end understanding from digital logic gates to application-level programming.
NeRF vs. 3D Gaussian Splatting: Comparative Analysis
Conducted comprehensive comparative analysis of state-of-the-art 3D reconstruction techniques: Mip-NeRF 360, Instant-NGP, and 3D Gaussian Splatting. Implemented end-to-end pipeline including COLMAP-based camera pose estimation on Tanks and Temples dataset. Evaluated models across three dimensions: rendering speed (FPS), view consistency (temporal stability), and detail preservation (PSNR, SSIM, LPIPS), achieving quantitative benchmarks for real-time vs. quality trade-offs in neural rendering.
Inverse Rendering with Cloud-Based GPU Infrastructure
Designed and deployed a cloud-based deep learning environment on AWS EC2 (g4dn.xlarge with NVIDIA T4 GPU) for inverse rendering research. Configured Deep Learning AMI with PyTorch/CUDA stack, implemented cost-efficient resource management strategies under student budget constraints, and established remote development workflows using VS Code Remote-SSH and Jupyter Lab with SSH tunneling for secure model training and experimentation.
Divide & Conquer: Long Video Understanding with LLMs
Developed a novel divide-and-conquer algorithm for long-form video understanding using Large Language Models. Implemented the 'conquer' phase logic that improved contextual understanding (CU) scores by 6%, enabling efficient summarization of 10+ minute videos. Designed chunking and hierarchical merging strategies to handle temporal dependencies across video segments.
Safe RL Framework with Control Barrier Functions
Researched and implemented a Shielded Learning framework integrating Control Barrier Functions (CBFs) with Soft Actor-Critic (SAC) and P2BPO optimization for Safe Reinforcement Learning. Ensured policy compliance with safety constraints while mitigating oscillation issues, demonstrating improved safety metrics in simulation environments.
Hierarchical RAG & Summarization Engine
Engineered a scalable RAG pipeline for long-context document retrieval using a hierarchical tree-based indexing mechanism (BFS/DFS traversal), reducing query latency by 40%. Integrated Fine-tuned Llama models with LoRA adapters to optimize summarization accuracy (ROUGE-L +15-25%) while cutting training costs by 80%.
JourneyGo: Group Travel Itinerary Recommender
Led a 5-person capstone team to build an intent-based travel recommendation web platform. Designed RESTful APIs and PostgreSQL schemas to support personalized itinerary generation. Integrated BERT models for user intent classification and recommendation ranking, achieving stable performance under load testing with 500+ real-world tourist attractions.
Slavic Languages Linguistic Search Engine
Architected a high-performance linguistic search engine supporting windowed proximity queries (e.g., 'A near B') via positional inverted indexing in PostgreSQL. Built an automated NLP pipeline using spaCy for complex morphological tokenization and lemmatization, bridging raw text data with structured linguistic research needs.