Megha Kalia

This project fine-tunes the Stable Diffusion XL (SDXL) model using DreamBooth combined with Low-Rank Adaptation (LoRA). Outputs include LoRA adapter weights ready for personalized image generation. LoRA enables fast and memory-efficient fine-tuning by updating only a small number of parameters.

A CrewAI multi-agent system that processes grocery receipts using Google Gemini's computer vision to extract and match items against a household inventory, then auto-generates a prioritized weekly shopping list based on what you have vs. what you need. \n Tech Stack Next.js 15 · Python/Flask · Google Gemini 2.5 Flash · CrewAI · Vercel + Render

A parameter-efficient fine-tuning project using 4-bit QLoRA on a T4 GPU. Includes full evaluation pipeline with loss tracking and perplexity comparison (91.7 → 74.4). Demonstrates practical, consumer-GPU LLM fine-tuning with quantitative and qualitative validation..

Self-supervised monocular pose estimation framework that jointly learns scene depth and relative camera motion directly from raw video sequences without requiring ground-truth annotations. By leveraging photometric and geometric consistency constraints within a PyTorch-based deep learning pipeline, the system provides a scalable, data-efficient solution for vision-driven motion estimation.

This project explores how much delay people can tolerate between what they see and what they feel in a Mixed Reality environment. Built in **Unity3D (C#)**, the system integrates an **HTC Vive** headset for immersive VR and a **3D Systems Touch** haptic device for force feedback. I programmatically introduced controlled delays between visual and tactile events to measure the point at which users begin to notice desynchronization. The goal is to better understand human tolerance to latency and provide practical insights for designing responsive robotics and haptic systems for metaverse and MR applications. (Done as part of Friedman Scholarship, New York Univerisity, USA)

Our method adopts a generation and segmentation strategy to learn a segmentation model with better generalization capability to domains that have no labelled data. The method leverages the availability of labelled data in a different domain. The generator does the domain translation from the labelled domain to the unlabeled domain and simultaneously, the segmentation model learns using the generated data while regularizing the generative model. Presented at MICCAI 2021 (Oral, Top 13%).

The project aims to provide enhanced vision to the surgeons for better surgical planning and accuracy. A patient’s MRI is deformably registered to the phantom. 3D mesh (pink) generated from MRI and the MRI plane is visible in the image. Transverse-plane of the MRI can is controlled using the surgical instrument. Notice how the MRI plane changes with the surgical instrument. At many locations tumor locations in red can be seen on the MRI plane.