Overview
- Creative Innovator: I excel at developing impactful tech solutions, from AR tools that enhance accessibility to blockchain-based applications that ensure security and transparency. My work consistently demonstrates a strong ability to identify problems and implement effective, cutting-edge solutions.
- Broad Technical Skillset: Proficient in multiple languages and frameworks, including Python, C#, Solidity, and more, with hands-on experience in machine learning models, VR/AR, and adaptive recommendation systems. My versatility as a developer enables me to adapt quickly and tackle a variety of technical challenges.
- Proven Leader & Effective Collaborator: I have led successful projects in both solo and team settings, earning awards at hackathons and delivering software solutions in global placements. I thrive under pressure, manage projects end-to-end, and communicate effectively to achieve high-quality results.
AR Handwriting Spellchecker with AI Powered Optical Character Recognition
Handwrite-AR is an augmented reality tool designed to assist individuals with dyslexia by providing real-time spell checking, punctuation correction, and text translation with 98% accuracy. This application overlays corrections on handwritten text, offers audible narration and dictionary definitions, and includes customizable color filters to enhance readability. Built using C# and OpenCV it integrates Google OCR, Microsoft Azure Spell Checker, and Google Translate APIs, processing base-64 images for efficient text detection and correction. Competing against over 500 participants in teams of four, Handwrite-AR won 1st place at Durham Hackathon 2023, along with the 3D Technology and AI Assistance category awards. The project showcases the potential of AI-driven accessibility tools, with real-world impact as one of our team members with dyslexia has already considered using it to aid his writing.
Efficient Image Classification with MobileNet (CIFAR-100)
I developed an optimized image classification model using a MobileNet-based Convolutional Neural Network (CNN) for CIFAR-100, focusing on efficiency and computational cost reduction. By leveraging depth-wise separable convolutions, the model significantly improved processing speed while maintaining accuracy. To enhance generalization and prevent overfitting, I implemented L1/L2 regularization, Adam optimization, and extensive data augmentation techniques, including horizontal flips, rotations, and color jittering. The final model achieved 60% test accuracy across 100 classes with just 102,916 parameters, demonstrating strong performance in a low-memory setting. This project highlights my expertise in efficient deep learning, model optimization, and image classification for resource-constrained environments.
Semantic Stock Prediction and Portfolio Optimisation Software
I developed a software platform that integrates semantic data with machine learning models to enhance stock price prediction. By leveraging BERT for sentiment analysis, the system analyzes financial news and reports, providing deeper market insights. To evaluate performance, I compared traditional deep learning models like LSTM and GRU against semantic-aware approaches, utilizing live financial APIs for real-time data processing. Additionally, I designed a portfolio optimization module that employs reinforcement learning and modern portfolio theory to maximize returns while minimizing risk, enabling more informed investment strategies. This project showcases my expertise in natural language processing, financial AI, and advanced machine learning techniques for dynamic market analysis.
Blockchain based Ticketing System
I developed a cutting-edge blockchain-based event ticketing system designed to revolutionise the way we buy and sell tickets. This application allows users to purchase, trade, and verify event tickets securely on the blockchain. Leveraging the power of smart contracts, each ticket is represented as a unique non-fungible token (NFT), ensuring authenticity and preventing fraudulent activities. The system supports a dynamic pricing model that adjusts based on demand, optimising profitability while minimising scalping activities. This project highlights the potential of decentralised technologies to enhance transparency, security, and user trust in the event ticketing industry.
To build this sophisticated system, I utilised an array of advanced tools and technologies. The smart contracts were developed using Solidity within the Visual Studio Code environment, and deployed on both Ethereum and Polygon networks to compare performance metrics. I employed the Hardhat development framework for its efficiency and comprehensive testing capabilities. Metamask served as the digital wallet interface, facilitating secure transactions between users and the blockchain. For the front-end, I used JavaScript with the Express framework, and integrated it with the blockchain via the Ethers.js library. Additionally, tools like Ganache and Truffle Suite were initially used for local development before transitioning to Hardhat for improved performance.
The achievement of this project is substantial. I successfully created a fully functional decentralised application (dApp) that mitigates common issues in current ticketing systems, such as scalping and hidden fees. By deploying the application on public blockchains, it offers enhanced security and transparency compared to traditional systems. The integration of dynamic pricing algorithms has proven to significantly increase event profitability, and the real-time updates ensure users are always informed about ticket availability. This project not only demonstrates my technical prowess in blockchain development but also showcases my ability to solve real-world problems using innovative technologies .
Reinforcement Learning
I developed an advanced reinforcement learning model to teach a 2D Bipedal Walker to navigate complex terrains in OpenAI Gym’s BipedalWalker-v3 environment, including the hardcore variant with obstacles like ladders and pitfalls. Using Python and PyTorch, I implemented a hybrid Twin Delayed Deep Deterministic Policy Gradient (TD3) with a Forward-Looking Actor (FORK) model, integrating predictive state modeling to enhance adaptive movement planning. This approach improved convergence by 45% compared to traditional TD3 algorithms, while Ornstein-Uhlenbeck noise optimization stabilized policy learning and reduced overestimation bias. The walker successfully mastered challenging terrains, demonstrating the effectiveness of TD3-FORK in high-dimensional control tasks. A project video showcases its optimal navigation performance, highlighting my expertise in reinforcement learning, algorithm optimization, and AI-driven decision-making for dynamic environments.
Conditional GANs for CIFAR-100 Image Generation
I implemented a Conditional Generative Adversarial Network (CGAN) for CIFAR-100 image generation, integrating class labels into both the generator and discriminator networks to enhance class-specific synthesis. The model follows a Deep Convolutional GAN (DCGAN) architecture, utilizing LeakyReLU activation, label concatenation, and adversarial training for improved stability and realism. Trained over 50,000 optimization steps, the 99M parameter model effectively generates diverse images while maintaining computational efficiency. The system achieved a Frechet Inception Distance (FID) of 85.94, balancing model size and generative quality. This project demonstrates my proficiency in deep generative modeling, adversarial training, and efficient image synthesis.
(Image is a placeholder. Due to being part of my degree, the image cannot be released to prevent plagiarism.)
Deep Learning Based Adaptive Film Recommended System
The goal of this project was to develop a hybrid recommendation system that mitigates the cold start problem by integrating collaborative filtering (CF) and content-based filtering (CBF) techniques. Using a film dataset with over 100,000 data points, the system prompts users to rate initial films, dynamically updating recommendations through a deep learning-driven convolutional neural network (CNN) in PyTorch. We meticulously cleaned the dataset to ensure accuracy and achieved an RMSE of 0.68 and a novelty score of 3.2, demonstrating strong recommendation quality. Evaluated using rigorous performance metrics, the system delivers personalized film suggestions with state-of-the-art AI techniques. Currently, it operates via the terminal, but plans are underway to integrate a graphical user interface (GUI) for an improved user experience.
VR Rendering System with Custom Physics Engine
I developed an advanced virtual and augmented reality system by enhancing the RenderPy-master Python repository, transforming static frame rendering into real-time 3D rendering using PyGame and NumPy. By implementing perspective projection and a homogeneous coordinate system, I improved depth perception and object transformations, while quaternion mathematics enabled smooth camera movements and dynamic scene adjustments. Additionally, I developed a comprehensive tracking system that fuses IMU accelerometer and gyroscope data for precise orientation and movement detection using quaternion transformations and dead reckoning. A physics engine incorporating collision detection, gravity, and air resistance enhanced realism, while an innovative Level of Detail (LoD) system dynamically adjusted object complexity, improving rendering efficiency by 32% without compromising visual fidelity. The project is showcased in a video demonstrating advanced physics simulations and smooth rotations, highlighting the seamless integration of real-time rendering, physics, and motion tracking for immersive VR experiences.
Smart Meal Planner and Food Waste Management System
As a finalist in the Durham university hackathon, I contributed significantly to the development of a this food waste management system. This innovative application allows users to input their ingredients and receive tailored recipe recommendations, prioritising items that are closest to their expiration dates to effectively minimise food waste. The system also suggests sustainable alternatives, such as donation options for unused items, thereby promoting ethical eating habits.
My role centered on backend development, where I architected and implemented robust REST APIs and set up a scalable database infrastructure. This involved crafting intricate SQL queries and ensuring seamless integration between the frontend and backend components. For the application we used Python and Flask to run the web sever and then a SQLite Database,
The project was both technically challenging and rewarding, requiring close collaboration with my teammates to resolve data formatting issues and enhance overall functionality. Participating in the hackathon provided invaluable experience in web development and teamwork, highlighting the importance of effective communication and time management in a high-stakes environment.
Bitcoin Autotrader
My Bitcoin auto trader is a sophisticated, Python-based application designed to optimise your cryptocurrency trading on Coinbase Pro. Utilising the Coinbase API, it employs advanced trading strategies, including dynamic moving stop losses, to minimise losses and maximise profits. As Bitcoin prices fluctuate, the stop loss levels adjust accordingly, ensuring your investments are protected. This dual functionality allows seamless trading between BTC and EUR, capitalising on small market movements with frequent trades within specified timeframes.
The intuitive Tkinter front end enhances user interaction, allowing you to customise stop loss percentages and re-buy amounts, directly affecting how your funds are managed on Coinbase Pro. Importantly, the application itself doesn't handle any funds; all transactions occur securely on Coinbase Pro, ensuring the safety of your assets. Users can monitor real-time market data and balances, execute trades, and configure their trading parameters through a user-friendly interface.
Additionally, the auto trader features a real-time graph displaying your profit trends, providing you with valuable insights into your trading performance. A multi-threaded design ensures the GUI remains responsive even while performing continuous market analysis and trading operations. Designed as a personal project it stands as a testament to the potential of automated trading to streamline and enhance your cryptocurrency investment strategies.
Unity Adventure Game based in Durham
D U Delivery is an immersive RPG game developed in Unity and C# with custom 2D graphics, designed to spark curiosity and problem-solving skills in children. Set in the historic city of Durham, players embark on an adventurous journey where they must solve riddles rooted in the city's rich history and mysterious rumours. Customise your character with various skins and enjoy real-time translation in multiple languages, making the game accessible to a diverse audience. Navigate iconic locations like the Durham Cathedral and the bustling marketplace using your keyboard and mouse. As you solve riddles and complete deliveries, be prepared to outsmart enemies with advanced AI path tracking. D U Delivery is more than just a game; it's an educational tool aimed at increasing tourism to Durham while fostering a love for history and enhancing problem-solving abilities in young players.