Hi, I'm Kritarth Dandapat

Lab Overview

The ESC group at the University at Buffalo delves into advanced research topics situated at the crossroads of hardware/architecture, operating systems, algorithms, human factors, and their applicability in healthcare, manufacturing, and cybersecurity domains. Our research undertakings encompass a broad spectrum, ranging from the development of hardware/software foundations to the construction of pragmatic computer systems.

Project Overview

Working on cutting-edge research in dental health technology, specifically focusing on AI-driven diagnostic tools and mobile applications for orthodontic assessment.

Key Responsibilities

• AI Model Development: Creating deep learning models for accurate orthodontic scoring using depth sensors and computer vision techniques
• Dental Health Application: Developing comprehensive AI-driven models that enhance diagnostic accuracy and improve user engagement
• Data Visualization: Implementing interactive features to visualize dental health progress over time
• Personalized Recommendations: Building systems that provide tailored health recommendations based on user data
• Cross-Platform Development: Building both mobile (React Native) and web (React) applications

Technologies Used

• React Native for mobile development
• React.js for web applications
• Python for AI/ML model development
• Computer Vision libraries
• Depth sensor integration

Impact

This research contributes to the advancement of digital health technologies, potentially improving access to dental care through innovative AI-powered diagnostic tools.

Lab Information

Learn more about the ESC Group: ESC Group Website

Role Overview

Serving as both a tutor and Peer-Assisted Learning (PAL) leader, helping students master statistical concepts through interactive sessions and personalized support.

Key Responsibilities

• Interactive Sessions: Conducting two 1-hour interactive sessions per week focused on Statistics
• Performance Improvement: Achieving 30% improvement in student understanding based on quiz performance and feedback
• Concept Reinforcement: Providing in-depth explanations and answering student queries during PAL sessions
• Assessment Support: Helping students prepare for exams and assignments
• Peer Mentoring: Serving as a role model and mentor for fellow students

Teaching Methods

• Interactive problem-solving sessions
• Real-world application examples
• Step-by-step concept breakdowns
• Group discussions and collaborative learning
• Individual consultation sessions

Impact

Helped numerous students improve their statistical understanding and academic performance, contributing to the overall success of the university's academic support program.

Organization Overview

Co-founded and led the University at Buffalo chapter of the National Student Data Corps, a nationwide initiative to promote data science education and collaboration among students.

Key Achievements

• Website Development: Led the complete development and launch of the NSDC website, significantly improving event coordination and member communication
• Team Leadership: Organized and facilitated collaborative website development, ensuring effective project planning and execution
• Event Management: Planned and executed various data science workshops and networking events
• Community Building: Established a strong network of data science enthusiasts at UB

Technical Contributions

• Full-stack website development
• Event management system implementation
• Member communication platform
• Resource sharing and documentation systems

Leadership Skills Developed

• Project management and coordination
• Team collaboration and communication
• Strategic planning and execution
• Community engagement and outreach

Impact

Successfully established a thriving data science community at UB, providing students with opportunities to learn, collaborate, and grow in the field of data science.

Project Overview

Developed a sophisticated deep learning system for accurate people counting in dense crowd scenarios, addressing the challenges of occlusion and varying crowd densities.

Technical Implementation

• CSRNet Architecture: Implemented the Crowd Scene Regression Network for robust crowd counting
• Occlusion Handling: Designed algorithms to handle partial visibility and overlapping individuals
• Multi-scale Processing: Utilized different resolution inputs to capture both local and global crowd patterns
• Real-time Processing: Optimized for live video stream analysis with minimal latency

Key Features

• Accurate counting in dense crowds (100+ people)
• Robust performance under varying lighting conditions
• Real-time processing capabilities
• Cross-platform compatibility
• Comprehensive evaluation metrics

Technologies Used

• PyTorch for deep learning framework
• OpenCV for image processing
• CSRNet architecture implementation
• CUDA for GPU acceleration
• Python for backend processing

Results & Impact

Achieved state-of-the-art performance on standard crowd counting datasets, with potential applications in security, retail analytics, and public safety monitoring.

Project Overview

Engineered a comprehensive emotion detection system that combines multiple deep learning architectures to accurately classify human emotions from facial expressions.

Architecture Design

• CNN Implementation: Custom convolutional neural network for feature extraction
• ResNet-34 Integration: Leveraged residual connections for improved gradient flow
• Vision Transformer (ViT): Applied attention mechanisms for global feature understanding
• Ensemble Approach: Combined predictions from multiple models for robustness

Performance Metrics

• Overall accuracy: 87.5%
• Support for 7 emotion categories
• Real-time inference capabilities
• Robust performance across different demographics

Technical Stack

• PyTorch for model development
• TensorFlow for comparison experiments
• OpenCV for face detection and preprocessing
• NumPy for numerical computations
• Matplotlib for visualization

Applications

Potential applications include mental health monitoring, human-computer interaction, marketing research, and accessibility tools for emotion recognition.

Project Overview

Developed an interactive pathfinding algorithm visualizer that demonstrates various search algorithms in real-time, with a focus on the A* algorithm and its optimization techniques.

Core Features

• Interactive Grid: Clickable grid system for setting start, end, and obstacle points
• Real-time Visualization: Step-by-step algorithm execution with visual feedback
• Multiple Algorithms: Implementation of A*, Dijkstra's, and BFS algorithms
• Customizable Parameters: Adjustable grid size, animation speed, and heuristic functions
• Path Optimization: Visual comparison of different pathfinding approaches

Technical Implementation

• Pygame for graphics and user interaction
• Custom A* algorithm with multiple heuristics
• Efficient data structures for path computation
• Modular design for easy algorithm addition
• Performance optimization for large grids

Educational Value

This project serves as an excellent educational tool for understanding pathfinding algorithms, data structures, and algorithm visualization techniques.

Future Enhancements

• Support for weighted graphs
• 3D visualization capabilities
• Algorithm performance comparison
• Export functionality for educational materials

Project Overview

Designed and implemented a comprehensive command-line shell specifically tailored for students, incorporating educational features, productivity tools, and gamification elements.

Key Features

• Custom Aliases: Pre-defined shortcuts for common student tasks
• Command History: Enhanced history with search and filtering capabilities
• Auto-completion: Intelligent command and file path completion
• Code Execution: Direct execution of programming assignments
• Interactive Tutorials: Built-in learning modules for shell commands
• Gamification: Achievement system and progress tracking

Educational Components

• Step-by-step command tutorials
• Interactive quizzes and challenges
• Progress tracking and achievements
• Contextual help and documentation
• Error handling with educational feedback

Technical Architecture

• Python for core shell functionality
• SQLite for user progress storage
• JSON for configuration management
• Modular design for easy feature addition
• Cross-platform compatibility

Impact

This project helps students develop essential command-line skills while making the learning process engaging and interactive, bridging the gap between theoretical knowledge and practical application.

Project Overview

Co-developed a comprehensive web application during the UB Hacking hackathon, focusing on classification tasks with a modern, responsive user interface and robust backend architecture.

Development Role

• Backend Architecture: Designed and implemented the server-side logic
• Database Design: Created efficient data models and relationships
• API Development: Built RESTful APIs for data processing
• Frontend Integration: Connected UI components with backend services

Technical Stack

• Frontend: React.js with Material-UI (MUI) components
• Backend: Node.js with Express.js framework
• Database: Firebase Firestore for real-time data
• Styling: CSS3 with responsive design principles
• Deployment: Firebase Hosting for web deployment

Key Features

• Responsive design for all device sizes
• Real-time data processing and visualization
• User authentication and data persistence
• Interactive classification interface
• Results export and sharing capabilities

Hackathon Experience

Successfully delivered a fully functional application within the hackathon timeframe, demonstrating strong teamwork, rapid prototyping skills, and technical problem-solving abilities.

Learning Outcomes

• Rapid development under time constraints
• Team collaboration and version control
• Full-stack development integration
• User experience design principles
• Cloud deployment and hosting

Project Overview

OralScan is an AI-powered platform that provides accessible, real-time oral health diagnostics and care recommendations, especially for seniors and underserved communities. The app leverages advanced AI models to detect oral health issues from smartphone images, offering instant feedback and personalized care guidance.

Key Features

• AI-driven diagnostics for oral health issues (cavities, gingivitis, ulcers, etc.)
• Mobile accessibility (Android & iOS) for at-home dental screening
• Real-time analysis and personalized care recommendations
• Guided camera system for easy image capture
• Web-based dashboard for admin and research monitoring
• Secure server for data storage and fairness assessment

Technologies Used

• React Native (mobile app)
• React (web dashboard)
• Python (AI/ML backend)
• YOLO, SAM, ResNet50-based CNN for image analysis
• Secure server and relational database

My Role

• Developed and tested the mobile and web app UI
• Integrated AI model APIs for real-time diagnostics
• Collaborated with a multidisciplinary team of researchers, doctors, and engineers
• Conducted user testing and contributed to app feature refinement

Impact

• Empowered seniors and rural/low-income populations with accessible dental care
• Enabled early detection and prevention of oral health issues
• Bridged healthcare gaps with cost-effective, real-time solutions

Learn More

oralscan.health

Lab Overview

The Peng Research Lab at the University at Buffalo focuses on working at the nexus of 🤖 (AI) + ⚛️ (Quantum Computing) →⚡ (Energy Systems) + 🧪 (Chemistry). The lab explores cutting-edge research at the intersection of artificial intelligence, quantum computing, and energy systems.

Research Focus

• AI Integration: Developing AI algorithms and machine learning models for complex system optimization
• Quantum Computing: Exploring quantum algorithms and their applications in energy systems
• Energy Systems: Researching sustainable energy solutions and optimization
• Chemical Applications: Applying computational methods to chemical and materials science

Key Responsibilities

• Research Contributions: Contributing to the lab's mission of advancing 🤖 + ⚛️ →⚡ + 🧪 technologies
• Collaboration: Working closely with graduate students and faculty on innovative research projects
• Algorithm Development: Implementing and testing novel computational approaches
• Data Analysis: Processing and analyzing research data using advanced computational methods

Technologies & Skills

• Machine Learning and AI algorithms
• Quantum computing frameworks
• Energy system modeling
• Computational chemistry tools
• Python and scientific computing libraries

Impact

This research contributes to the advancement of cutting-edge technologies that bridge AI, quantum computing, and energy systems, potentially leading to breakthroughs in sustainable energy and computational chemistry.

Lab Information

Learn more about the Peng Research Lab: Peng Research Lab Website

Project Overview

The Marine Guardian project aims to develop a model for detecting ships in satellite imagery, crucial for maritime security, environmental monitoring, and naval operations. The goal was to establish baseline performance using existing models and then explore first principles of computer vision for faster, real-time detection.

Technical Challenge

The project addressed the trade-off between accuracy and speed in ship detection. While well-defined CNNs provide good accuracy, they lack the speed needed for live monitoring. The approach was inspired by face detection, where Haar filters provide faster detection than CNNs.

Dataset & Methodology

• Dataset: Ships in Satellite Imagery from Kaggle (1000 ship images, 3000 no-ship images)
• Challenge: Dataset labeling inconsistencies with partial ships classified as "no-ship"
• Approach: Combined transfer learning with first principles of computer vision

Technical Implementation

• Baseline Models: Basic CNN, EfficientNetB0, MobileNetV2
• Transfer Learning: Used pre-trained EfficientNet as encoder with custom decoder
• Unsupervised Learning: ResNet50 + KMeans clustering for dataset validation
• First Principles: Roundness-based classification using geometric properties

Key Results

• MobileNetV2: 98.72% accuracy, 99.31% validation accuracy (best performance)
• EfficientNetB0: 74.25% accuracy (affected by dataset issues)
• Roundness Algorithm: Detected 2234 ship images (close to expected 2000)
• Speed Optimization: Fast detection using geometric properties for real-time monitoring

Technical Innovations

• Roundness Calculation: Developed algorithm using area-to-perimeter ratio for shape classification
• Dataset Validation: Uncovered and addressed labeling inconsistencies
• Hybrid Approach: Combined deep learning with traditional computer vision techniques
• Real-time Processing: Optimized for live satellite imagery monitoring

Technologies Used

• Python with OpenCV for image processing
• TensorFlow/Keras for deep learning models
• EfficientNet, MobileNetV2, ResNet50 architectures
• KMeans clustering for unsupervised learning
• Geometric algorithms for roundness calculation
• Google Colab for GPU-accelerated training

Impact & Applications

This project demonstrates the importance of dataset validation and the effectiveness of combining modern deep learning with traditional computer vision techniques. The roundness-based approach provides a fast, efficient method for ship detection that can be deployed in real-time maritime monitoring systems.

Lessons Learned

• Thorough dataset validation is crucial before model development
• Simple geometric solutions can outperform complex neural networks in specific applications
• Documentation and organization of experimental results is essential
• First principles thinking can lead to unexpected but effective solutions

Dataset Reference

Ships in Satellite Imagery Dataset