Sushovan Adhikari
Machine Learning Engineer
Building intelligent systems that solve real-world problems through the intersection of data science and software engineering.
View My Work
Building intelligent systems that solve real-world problems through the intersection of data science and software engineering.
View My WorkI'm a passionate Machine Learning Engineer with a Master's degree in Data Science and a strong background in software development. My unique combination of theoretical knowledge and practical engineering skills allows me to build scalable, production-ready ML systems that deliver real business value.
I specialize in end-to-end machine learning pipelines, from data preprocessing and model development to deployment and monitoring in production environments. My experience spans various domains including time series analysis, reinforcement learning, and cybersecurity applications.
Master's thesis research investigating the threat of AI-generated phishing emails. Trained an LSTM-CNN model and fine-tuned BERT, achieving 98% and 99% accuracy on human-crafted email datasets. When tested on 711 AI-generated phishing emails, both models misclassified at least 10% as legitimate, revealing critical security risks. Fine-tuning showed BERT (99.48% accuracy) significantly outperformed LSTM-CNN (48.61%). This work highlights the urgent need for stronger cyber defenses against AI-generated threats.
Built a full-stack web application using Django REST API and React TypeScript that leverages AI to detect phishing and spam emails in real-time. Integrated Gmail OAuth 2.0 for secure email access, implemented JWT authentication, and created an intuitive dashboard for threat analysis with confidence scoring. Features include advanced email parsing, CSV batch processing, and a responsive interface for comprehensive email security analysis.
Statistical analysis of COVID-19 temporal dynamics using cross-correlation techniques and linear regression modeling to quantify case-fatality relationships across three pandemic waves, revealing significant improvements in healthcare response patterns and mortality reduction from 3.85% to 1.28% over time.
Developed a reinforcement learning solution for the OpenAI Gym Mountain Car environment using Q-learning algorithms with 20x20 state space discretization. Implemented epsilon-greedy exploration strategies achieving 85%+ success rates across 5,000 training episodes, with optimized hyperparameters (learning rate: 0.1, discount factor: 0.99) and extended episode limits of 500 steps.