Portfolio
From multi-agent RL to NASA citizen science - things I've built and researched.
Interactive 3D globe visualising global statistics sourced from Our World in Data - covering population, COâ‚‚ emissions, HDI, deforestation, and energy consumption. Countries are colour-coded by metric and clickable, opening a side panel with per-country details. Includes an animated timeline that scrubs through historical data with live colour transitions, showing how the world has changed over time.
Browser-based dashboard making 26 years of Formula 1 data explorable in one place. Past seasons (2000–2025) load from Supabase in under a second; the current season fetches live from the Jolpica F1 API. Features race results, qualifying times, championship standings, grid-to-finish slope charts, and six interactive analytics charts - no account, no paywall.
Portfolio allocation system using Multi-Agent Reinforcement Learning. Leading development of individual agent models and a meta-model that coordinates across agents to optimise decisions. Designing financial data pipelines covering stocks, crypto, and ETFs from 2010 to present. Integrated with Alpaca for live paper trading simulation.
Contributed to a NASA-supported citizen science initiative analysing solar eclipse imagery at scale. Developed a specialised image alignment tool using OpenCV for large-scale astronomical image processing. Collaborated across a multidisciplinary team on data preparation and scientific workflows.
Open-source Python GUI for the Chip Test Board at Paul Scherrer Institute (PSI). Enables researchers to configure and control X-ray and photon detector hardware with live signal visualisation. Presented as a poster at NoBugs Conference 2022, Switzerland - publicly used by researchers at PSI.
Pre-processed 10K+ mango leaf images from Dhaka University Agriculture Department using OpenCV, with augmentation (resize, normalise, rotate) for generalisation. Built and trained a custom CNN achieving ~97% accuracy on validation data. Deployed via Flask and Hugging Face Spaces with an accessible Streamlit UI - built for use in remote farming communities.
Collected and processed 1.6M+ unstructured healthcare reviews - cleaning, tokenisation, lemmatisation, and TF-IDF vectorisation. Built scalable GCP workflows for data ingestion and model training. Deployed a real-time NLP pipeline and Streamlit UI with live sentiment tracking and visualisation dashboards.
Web app that condenses long articles into concise summaries using five models - LSTM (RNN), Logistic Regression, Linear Regression, and two Decision Tree variants - trained on the CNN/DailyMail dataset. Users choose the model and summary length. Personally responsible for the LSTM model, output merging, and full deployment.
Intelligent recommendation engine combining three ML approaches - KNN, Content-Based Filtering, and User-Based Collaborative Filtering - across a dataset of 500K+ entries. Fuses outputs from all three models to generate 5 curated, bias-reduced recommendations per user based on real-time input.
Full-stack Android medical companion app providing instant access to health information and emergency resources. Features medical condition search with descriptions, symptoms, prevention, and cures. Includes location-based hospital, doctor, and ambulance finder. Built a community-driven blood bank where users can find or donate blood by type.