projects

Research Projects

Image2Blocks: Reconstructing Images in 3D using Cuboids

Goal: Monocular 3D scene understanding using primitive-based representation

• Formulated monocular image understanding as reconstructing scenes using 3D cuboid primitives rather than only 2D attributes
• Combined monocular depth and point-cloud reconstruction to infer layout, free space, and occlusions in globally consistent 3D scene models
• Enables intuitive 3D scene editing and manipulation

Text-Guided Diverse Image-To-Image Translation

Goal: Controllable and diverse image transformation using natural language

• Developed a model for text-guided image-to-image translation that preserves the main object while applying text-specified transformations
• Leveraged pre-trained language embeddings and conditional image synthesis to produce realistic, diverse outputs aligned with the text prompt
• Supports creative applications like style transfer, object modification, and scene editing through natural language

Industrial Projects

IoT2Cloud Platform

Role: Co-designer and ML Lead

• Co-designed and delivered IoT2Cloud platform for web, iOS, and Android
• Integrated LLM-based natural language to UI-component generation
• Built custom evaluation framework and curated UI component dataset
• Successfully deployed across multiple platforms with real-world usage

Offline Voice Assistant

Role: Lead ML Specialist

• Designed and deployed offline voice assistant models on embedded boards and smartphones
• Enabled on-device speech recognition and intent understanding without internet connectivity
• Optimized for resource-constrained embedded environments
• Achieved production-quality accuracy and latency

Vital Sign Monitoring System

Role: Team Leader

• Led development of transformer-based models for blood pressure estimation and vital-sign forecasting
• Processed PPG, ECG, and other vital signals for medical applications
• Deployed models meeting medical device accuracy requirements
• Optimized for real-time inference on edge devices

Maya i-Box - Intelligent Embroidery System

Role: A.I. Developer and Product Manager

• Developed and launched Maya i-Box, an intelligent embroidery design system
• Converts photographs and touchscreen sketches into optimized DST stitch files
• Designed complete computer vision and image-processing pipeline in C++/Qt
• Implemented computational geometry and graph-based optimization algorithms
• Successfully deployed in industrial production environments