Highlights

Python, Scikit, Keras, NumPy, Matplotlib

• Benchmarked three classifier families (RBF-SVM+PCA, MLP, and a VGG-style CNN) on the PathMNIST histopathology dataset (40k images across 9 tissue classes), tuning at least three hyper-parameters per model via cross-validation and validation-set search.
• Built a small CNN reaching 85.5% test accuracy and 0.851 macro-F1, beating the kernel SVM by ~20 and the MLP by ~32 accuracy points despite having the fewest parameters.
• Diagnosed model behaviour beyond headline accuracy using per-class F1, confusion matrices, and precision/recall trade-offs, showing that the CNN's spatial inductive bias (not raw capacity) drives the gain on texture-defined classes.

Next.js, TypeScript, Firebase, AWS OpenSearch, Gemini API

• Designed a legal search platform for Nepali lawyers, integrating AWS OpenSearch to index and query 360k+ scraped legal documents from official sources, achieving 60+ active users daily.
• Built a Next.js web application with Firebase authentication, tiered subscriptions, payment processing, and child-account management for law firms; deployed to Vercel with automated CI/CD.
• Implemented semantic search and chat-with-document functionality by integrating the Gemini API to process document context and generate conversational answers.

Next.js, TypeScript, Supabase, Tailwind, Vercel

• Engineered a full-stack Content Management System for a community school to digitize communications and operations for 300+ students and staff.
• Developed a secure admin dashboard with role-based authentication for managing school activities, official documents, and class statistics.
• Implemented a built-in mailing system to streamline internal communication between 12 administrators.
• Automated deployment via Vercel's CI/CD pipeline, ensuring seamless updates and high availability of the platform.

C++, SFML, Verlet Integration

• Developed a cloth simulation system using Verlet Integration in C++ and SFML.

React, TypeScript, Supabase, Cloudflare Pages, GitHub Pages

• Deployed on Cloudflare Pages with a GitHub Pages-based ticketing system for task management.
• Set up TypeScript React stack with Supabase backend, and configured ESLint for code quality.
• Built CI/CD pipeline to ensure successful linting and Cloudflare compilation pre-merge.
• Led code reviews and added features to enhance project functionality.

Python, JavaScript, ArcGIS, Scikit, Firebase, Supervised ML

• Selected as a top 40 global finalist by NASA among 700+ teams from 152 countries for the project.
• Developed a wildfire navigation system using crowdsourced reports, MODIS data, and an AI model (Random Forest) to predict fire risks with 82% accuracy.
• Integrated Google Earth Engine API to gather climate data, clipped to Nepal’s forests, and visualized risks with ArcGIS for real-time monitoring.