Things I've built.

DocDjinn

General-purpose framework built on vision-language models and diffusion-based handwriting models for controllable generation of full-page synthetic documents, along with their annotations, to train models across diverse downstream document understanding tasks

• #AI
• #Research

AtriaX

Web-based document AI platform built on Atria for dataset and model management, training, evaluation, and explanation visualization

• #AI
• #Development
• #MLOps

TorchXAI

Explainable AI toolkit with 20+ explainability metrics and 11+ multi-output explainers, enabling fast, scalable evaluation of interpretability in multimodal deep learning models

• #AI
• #Development
• #Research

DocVCE

Explainability method for generating visual counterfactuals via guided latent diffusion to interpret the decisions of document classifiers.

• #AI
• #Research

DP-DocLDM

Differentially private document image dataset generation using layout-conditioned latent diffusion models.

• #AI
• #Research

DocPrivacy

Privacy-preserving multimodal document information extraction through differential privacy, federated learning, and encrypted inference.

• #AI
• #Research

DocXClassifier

Augments CNN-based classifiers with pyramid-attention modules to introduce inherent interpretability.

• #AI
• #Research

ColDBin

An approach to document image binarization through cold diffusion.

• #AI
• #Computer vision
• #Research

DocAL

A complete active-learning toolkit for deep learning classification tasks.

• #AI
• #Development

DocRobustness

A toolkit for generating and evaluating corruption-robustness benchmarks for document images.

• #AI
• #Computer vision
• #Research

DocXAI

A toolkit for explainable AI in document image classification.

• #AI
• #Research

DeepStream Media

An NVIDIA DeepStream pipeline for real-time detection and tracking of individuals in news media.

• #AI
• #Development
• #Computer vision

DeepStream Retail App

An NVIDIA DeepStream pipeline for real-time traffic tracking and analytics in retail stores.

• #AI
• #Development
• #Computer vision

Generalized Whole Body Motion Architecture for a Humanoid Robot NAO

Master's Thesis

A generalized whole-body motion architecture enabling real-time dynamic motion tasks on NAO humanoid robots.

• #Robotics

Dynamically Stable and Impact-Controlled Humanoid Kick

Bachelor's Thesis

A momentum-control-based omni-directional kick engine for the NAO humanoid robot, achieving approximately 90% average kicking accuracy.

• #Robotics

Mavros / MoveIt

Integrates MoveIt motion planning with MAVROS-based aerial robots in ROS.

• #Robotics
• #Development

RosGym

Integrates Gym with ROS to train reinforcement-learning agents in ROS and AirSim environments.

• #Robotics
• #AI
• #Development

Irl2DRos

A ROS navigation package for 2D mobile robots, focused on the Pioneer 3-AT platform.

• #Robotics
• #Development

NAO Autonomous Robotics Stack · RoboCup SPL

Robotics stack integrating vision, motion planning, state-estimation/localization, and multi-robot coordination for fully autonomous RoboCup SPL gameplay.

• #Robotics
• #Development

UAV Exploration

Frontier-based exploration for navigating unknown environments and generating octomaps.

• #Robotics
• #Development

ORB-SLAM

A custom ROS implementation of ORB-SLAM for monocular and RGB-D cameras with bundle adjustment.

• #Robotics
• #Computer vision
• #Development

VFH3D+

An octomap-based VFH+ algorithm in 3D for local planning of UAVs and other robots operating in 3D.

• #Robotics
• #Development

Fusion of IMU and Vision Observations for State Estimation of a Quadrotor

A ROS and Gazebo case study fusing monocular optical flow with IMU measurements for quadrotor position and orientation estimation.

• #Robotics
• #Computer vision