Things I've
built.

The Problem

Power grids fail. When faults in medium-voltage transmission lines go undetected or are misclassified, the result is outages, equipment damage, and costly downtime. Most existing systems rely on simple threshold-based alerts that generate false positives and miss subtle faults entirely.

What I Built

An end-to-end AI system that detects and classifies faults in transmission lines before they escalate. Modeled transmission line behavior across multiple fault conditions (line-to-ground, line-to-line, three-phase faults) using MATLAB and Simulink.

Extracted features from voltage and current waveforms, then trained Random Forest and SVM classifiers on that data. The models learned to distinguish fault types with high accuracy.

Wrapped everything in a Streamlit dashboard so engineers can see real-time 3-phase waveform visualization and fault predictions without needing to understand the ML pipeline underneath.

What I Learned

Feature engineering matters more than model complexity. Spent more time on extracting the right signal features than on tuning hyperparameters. Also learned that dashboards need to be designed for the person operating them, not the person who built the model.

The Problem

NSUT students needed a trusted, campus-specific platform to buy and sell textbooks, notes, and electronics with other students. Generic platforms weren't built for this context, and campus bulletin boards were a mess.

What I Built

A full-stack marketplace using the MERN stack. Students can list items, browse by category, reserve products, and message sellers. Authentication is JWT-based. The UI is responsive across all devices.

Used Redux Toolkit for state management to keep the frontend predictable. Real-time listing updates mean students see current availability. Implemented optimistic locking to prevent race conditions on product reservations.

Built an admin dashboard where moderators can review flagged listings, manage user accounts, and view platform analytics. Deployed the frontend on Vercel.

What I Learned

Managing concurrent state in a marketplace is genuinely hard. Concurrency, caching, and optimistic UI require thinking about edge cases from day one. Redux felt like overkill at first, turned out to be worth it at scale.

The Problem

Needed a reliable, low-cost device that automatically plays devotional music at scheduled times every day. Available solutions were either too expensive, too complex, or didn't allow enough customization for specific timing requirements.

What I Built

An Arduino-based music player using RTC (real-time clock) module synchronization. The device plays pre-stored tracks at configured daily times without any manual input.

Added Bluetooth control using the HC-06 module so users can update schedules, adjust volume, or skip tracks remotely from a mobile terminal app. No need to physically interact with the device.

Designed the PCB layout in KiCad for a cleaner, more compact and repeatable build. Low-power design means it runs reliably without overheating or requiring frequent maintenance.

Broader Use Cases

The design translates directly to school bell systems, automated public announcements, or any scheduled audio playback in low-power embedded environments. The core pattern is reusable.

What I Built

A conversational AI chatbot using OpenAI's GPT API with a Streamlit-based interface. Context-aware across turns, clean error handling, responsive UI. First project where I shipped a live ML application to actual users.

What I Learned

Integrating third-party APIs cleanly takes more thought than the API call itself. Rate limits, error states, and session context all need upfront consideration. Learned that UX for AI applications is its own discipline.

This project set the foundation for everything ML-related I've built since. Starting simple and getting something in front of people is better than over-engineering before launch.