Atharva Gundawar

Superior Computer Chess with Model Predictive Control, Reinforcement Learning, and Rollout

arXiv preprint arXiv:2409.06477

Robust Planning with LLM-Modulo Framework: Case Study in Travel Planning

arXiv preprint arXiv:2405.20625

LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks

arXiv preprint arXiv:2402.01817

On the Performance of new Higher Order Transformation Functions for Highly Efficient Dense Layers

Neural Processing Letters 55 (8), 10655-10668

Enhanced Dense Layers Using a Quadratic Transformation Function

Gundawar, A., & Lodha, S. (2022, December). In Proceedings (pp. 3-15). Cham: Springer Nature Switzerland.

SQL injection and its detection using machine learning algorithms and BERT

Lodha, S., & Gundawar, A. (2022, November). In Proceedings (pp. 3-16). Cham: Springer Nature Switzerland.

Research Assistant

• Improving Reasoning capacities of LLMs through neuro-symbolic frameworks

• Leveraging Foundational Models as a general solver for complete information environments

• Interactive Uncertainty Reduction for Efficient Vision-Language Spatiotemporal Navigation

Computer Vision Intern

• I devloped the foundation of a 3D shot-suggestion model over the exsiting 2D shot-suggestion model using NeRF (Neural Radiance Fields) and Pix2Vox framework.

• The framework to enable dynamic selection of style or content percentages, resulting in real-time similarity scoring, aligning with product design and artificial intelligence principles.

Research Intern

• Enhanced neural network performance by 4% through the implementation of higher-order transformation functions, contributing to machine learning algorithm development.

• Developed a framework to enable dynamic selection of style or content percentages, resulting in real-time similarity scoring, aligning with product design and artificial intelligence principles.

Machine Learning Intern

• Developed agile AI pipelines using Python, TensorFlow, and PyTorch to process 1000+ images per second for background and watermark removal, improving processing speed and efficiency.

• Optimized Neo4j database architecture, reducing storage space by 40% through the development of efficient data transfer and analysis pipelines.

Software Development Intern

Engineered an analytics system utilizing Graph Neural Networks (GNNs) to evaluate GitHub profiles across multiple dimensions—commit frequency, programming languages, contribution patterns, and peer reviews—generating a comprehensive profile score for applicant ranking on the platform.

• Crafted a state-of-the-art resume parsing engine, employing a combination of DistilBERT for semantic analysis and Transformer OCR (TrOCR) for data extraction, enabling the auto-population of applicant details across diverse resume formats.

Machine Learning Intern

• Analyzed Data for multiple UK-based Universities via EduMates and returned predictions and demographics through the data collected by the platform. Data included chats, user data, user behavior, and other undisclosable features.

Computer Vision Intern

• Developed an AI system for security feed analysis using YOLO for object detection and LSTM networks for temporal pattern recognition, enabling real-time anomaly detection. Leveraged convolutional and LSTM networks to process spatial-temporal data efficiently.

• Led the development of a VR-oriented Computer Vision system, integrating 3D CNNs for depth sensing and GANs for texture synthesis. Employed Pix2Vox and NeRF for 3D scene reconstruction from 2D inputs, enhancing VR immersion and realism for a client's trial.

Arizona State University

GPA: 4.0

ALGORITHMIC TRADING BOT

• Implemented a time sequence model which traded crypto which used technical indicators and market setntiments as the state space for the State Space models. Real profits were made with an average monthly ROI of 1%.

STYLE SIMILARITY

• Defined a new loss function called the Exponential contrastive loss function to calculate the style difference between two images. To test and demonstrate the working of the loss function, a style transfer model was used, with different levels of style transfer used to calculate the efficiency of the function.

ZERO-SHOT, ONE-SHOT, FEW-SHOT LEARNING FRAMEWORK

• Built a framework for easy training of few-shot models which deal with images. This framework allowed the user to use transfer learning on pre-trained siamese networks.

COMICCALL

• IterLUNet in a Single-image dehazing framework was used to make a real time image compression decompression system. This was then used to build a low-bandwidth video call application

COMMITMAN

• Implemented a Version control system from scratch in Python, which provides a robust CLI and a user-friendly GUI. The VCS can initialize a new repo, commit a new version, roll back to a previous version, and support branching and merges. Moreover, this works over the pre-existing .gitignore file for an easy shift from Git.

Anti-Face-Spoofing

• Implementation based on https://paperswithcode.com/dataset/replay-attack

Face-Generation

• DCGAN on CelebFaces Attributes Dataset (CelebA)

Face Recognition with Liveliness Detection

• Attendance with liveliness detection

Huber and Friends

• Implementation of Huber and Pseudo-Huber loss for regression and its Variant for classification