An evaluated, citation-grounded RAG system for Ontario auto insurance regulations. Features dense + BM25 hybrid retrieval fused by Reciprocal Rank Fusion (RRF), a refusal retry ladder boosting correctness to 80.7%, and strict provision-level citation grounding evaluated against a 62-item hand-labeled golden dataset across 3,896 legal document chunks.

An agent-based digital twin of Toronto designed for demand-informed renewable energy siting. Features a FastAPI simulation engine, an optimizer, and a planner connected to a React-based spatial map UI.

An AI-powered grief support platform that recreates a lost loved one's messaging style and voice from uploaded conversation histories. Features real-time voice synthesis and a coping-mode support toggle.

An NLP-based text classification project that distinguishes human-written content from AI-generated text. It compares two detection approaches: a classical machine learning pipeline using TF-IDF feature extraction with a linear classifier, and Transformer-based encoder models such as fine-tuned DistilBERT and RoBERTa.

Built an AI-powered email assistant using LangGraph and Groq that automates Gmail workflows through a natural-language chat interface. Developed tool-calling agents to search, read, classify emails, and generate Gmail draft replies while maintaining a draft-only safety model. Implemented end-to-end workflow orchestration to process unread emails through classification and response generation pipelines.

An AI computer vision project fine-tuned to classify various food images efficiently. Evaluated model precision using confusion matrices and tracked training history to optimize convergence and improve prediction accuracy on a custom dataset.

Built a full-stack parking locator for Toronto drivers that combines map-first search, nearby lot discovery, and route-aware navigation in one flow. Integrated Google Maps for location context and built a reliable backend layer to support stable requests as traffic and features scale.

An intelligent scheduling application developed during the Google AI hackathon to help users actively 'pilot' their week. It leverages modern web technologies and AI integrations to optimize user routines and intuitively manage weekly tasks.

A collection of robust Python scripts that retrieve comprehensive YouTube channel, playlist, and video metadata via the YouTube V3 API. The pipeline processes this data and synchronizes the results with Google Sheets, architected to feed downstream GenAI and research applications.

A web application that takes a YouTube video URL, fetches its transcript, and uses an AI LLM to generate a concise, markdown-formatted summary. Includes advanced UI features like playable video embedding, clickable UI elements for timestamps, and dynamic summary exports.

Built an interactive web app that visualizes meteor-impact scenarios using real NASA and USGS datasets. Developed a React and TypeScript interface with a 3D globe so users can inspect impact zones, compare simulated outcomes, and understand potential response strategies.

Built a research-oriented captioning system that combines computer-vision feature extraction with sequence-based language modeling to generate richer image descriptions. Explored multi-label prediction ideas to improve object coverage, semantic detail, and caption quality.

An offline, AI-driven solution combining visual analysis from an ESP32-S3 camera with temperature and smoke sensor data to detect fires in urban settings. A lightweight YOLOv8 model hosted on a Flask server processes images and cross-validates with hardware sensors to minimize false alarms.

Developed a secure Django-based file sharing application where users can upload, manage, and distribute files through time-limited links. Focused on authentication, access control, and straightforward sharing workflows to ensure teams can collaborate securely.

A command-based pathfinding game that utilizes data structures and algorithms to compute optimal paths and suggest new coordinates for navigating a maze towards the exit. Demonstrates strong understanding of foundational computer science principles through an interactive simulation.

A lightweight C++ logging library designed for embedded systems that streams log entries—including severity, file, function, and line information—to a central UDP server. The solution scales from a single device to multiple machines, featuring runtime log-level filtering.

A straightforward full-stack Todo application that lets users add, view, and delete notes. The backend runs on Node.js with an Express server, uses Zod for robust data validation, and stores data in a cloud-hosted MongoDB database via Mongoose, serving a modern React front-end.

Built a secure daily-post platform with authenticated CRUD APIs for predictable state handling. Implemented JWT-based route protection, modular Express endpoints, and clean data flow so content operations remain reliable and maintainable as features evolve.