Vinayak Chaturvedi

Machine Learning Engineer with background in independent research having core skills in NLP (Natural Language Processing) using traditional black box ML methods and deep learning with documented history of discovering methods to intelligently use data and enhance user experience. 4+ years of experience creating machine learning models, performing data analysis, re-training systems, and transforming data science prototypes to production-grade solutions. Consistently employs statistical methods and research techniques to provide novel solutions for the problem at hand.

• Malware Analysis by combining dynamic tracing and machine learning: Detected and removed malware in Windows portable executable (PE) files through utilization of machine learning techniques for in-depth dynamic analysis using Cuckoo Sandbox to extract relevant features (API calls and DLL counts). Executed malware samples inside a sandbox environment and study the API calls, memory, and associated registry keys. Further, the study was extended by employing state-of-the-art machine learning and deep learning methods such as LSTM, CNN, GRU, etc. to identify the malware samples. Used Linux as host system to run the cuckoo sandbox.
• Utilized the concept of transfer learning involving multiple models and techniques, such as state-of-the-art transformer networks (BERT, DeBERTa, RoBERTa, etc.), neural network architectures, feature-based models, subset-selection, textual entailment, and data augmentation to coordinate resolution of BoolQ task from SuperGLUE benchmark.
• Developed and deployed deep neural networks, including CNN, LSTM, and Bi-LSTM and Conditional Random Fields in Python via Keras and TensorFlow to forecast winning hero sequence in drafting phase of DoTA 2, massive multiplayer online battle arena game.
• Built "Hoax-Buster" - A deep learning (LSTM) based news misinformation classifier which separates legit news articles from hoax. The classifier was trained after data preprocessing stage involving stemming, lemmatization and stop word removal. Trained on New York Times dataset, this model performs with 99% accuracy and adapts well to new news articles on the web. This project was a part of New Hampshire Hackathon (NH Hack 2021) event.
• Formulated a 2D simulation car driving environment using Pygame and deep learning nets such as HandSegNet, and PoseNet to use hand gestures to drive the car.
• Accurate prediction of bird species distribution through utilization of R despite having presence-only bird data.

• Received scholarship from National Science Foundation (NSF) to be a part of Robotics Conference in Washington D.C. in 2020.