Hari Patel

Accomplished software engineer with extensive expertise in Python, Java, JavaScript, and TypeScript, complemented by proficiency in frameworks such as ReactJS, Express.js, and Node.js. Demonstrates a strong foundation in machine learning and deep learning technologies, including TensorFlow, Keras, PyTorch, and Scikit-learn. Adept at developing robust solutions using large language models and natural language processing techniques. Skilled in utilising Docker for containerisation and Git for version control. Committed to leveraging technical skills to drive innovation in software development while pursuing opportunities to advance expertise in reinforcement learning.

• IBM Software Engineering Module (Second-Year), 2024-10-01, 2025-03-31, https://github.com/COMP2281/software-engineering-group24-25-10, Collaborated in a team project with IBM to build a cross-platform application for automating PowerPoint generation and tracking engagement metrics in academic settings., Collaborated across frontend (ReactJS) and backend (Python, BERT) teams, applying knowledge of both to integrate FastAPI middleware and ensure smooth, reliable JSON data transfer between client and server., Delivered a functional, production-ready system with a complete deployment guide and documentation, ensuring the application could be reliably set up, scaled, and maintained as per module requirements.
• AI trimodal dataset analysis, Used logistic regression (one-vs-all with one-hot encoding) to classify a multidimensional trimodal dataset (text, image, EEG)., Performed data exploration, outlier removal (Isolation Forest), and PCA for dimensionality reduction (95% variance), achieving 0.5648 accuracy, 0.62 precision, 0.55 F1, and 0.56 recall (weighted)., https://colab.research.google.com/drive/1HnUmp5_yvMdXwpvP01HpqK2TT0OJzOl5?usp=sharing
• Deep Learning Models, https://github.com/Hari-P-22121, Developed a transformer-based classifier using a CNN encoder for the JetNet dataset, achieving 75.6% accuracy and 0.755 macro F1 across 5 jet classes (gluon, quark, top, W, Z)., Built a 3D CNN + Multi-Head Attention model in Keras to classify Parkinson's Disease from 3D brain MRI (PPMI dataset), achieving 0.74 accuracy, 0.70 F1 score, and 0.78 ROC-AUC.