Makungumixo Ndlovu

Geometry Area Calculator

Calculate areas of rectangles, triangles, and circles with an interactive interface.

Sudoku Solver

An interactive browser-based Sudoku puzzle solver that uses backtracking to solve any puzzle you load.

Prime Number Visualizer

Explore prime numbers up to any value using the Sieve of Eratosthenes. A clean and colorful way to visualize mathematics in action!

Math Facts Finder

Type in any number to discover an interesting trivia or math fact about it — fetched live from the Numbers API!

Tintswalo’s Poultry Delivery Tracker

A real-time delivery route optimizer for my broiler chicken business using Leaflet.js. It helps minimize fuel cost and delivery time by calculating the most efficient route between multiple customer addresses. Includes admin mode for route updates, cost estimation, and live mapping.

Fine-Tuning Low-Resource Translation Models Using Data Augmentation

This research focuses on improving Xitsonga-English translation by applying advanced data augmentation techniques to enhance low-resource models.

Google Buzz Legacy System (Software Engineering)

This academic project reimagines Google Buzz — a social networking and microblogging platform — by detailing both functional and non-functional requirements alongside an in-depth architectural design. It addresses performance, security, reliability, and usability concerns, illustrated with specific use cases and deployment models.

Key Highlights Include:

• Clearly defined functional requirements (user registration, profile management, privacy controls, etc.)
• Identification and quantification of non-functional requirements, such as latency goals and security measures
• System architecture combining Microservices, 3-Tier, and MVC patterns for robust scalability and maintainability
• Detailed use-case diagrams and interaction flows
• Deployment model leveraging Google Cloud Platform services

View Paper (PDF)

Hybrid Learning in Neural Networks for Almond Classification

This study explores how hybrid learning techniques enhance neural network performance for classifying different almond types (e.g., Mamra, Sanora, Regular). By combining two powerful gradient-based algorithms—Rprop and Adam—and conducting a grid search on key hyperparameters like batch size and epochs, the research demonstrates how smaller batch sizes (e.g., 16) and more epochs (up to 200) yield significantly better classification accuracy. The hybrid approach not only improves classification accuracy but also reduces overfitting and enhances generalization.

Key Highlights Include:

• Comparison of Rprop vs. Adam and a combined (hybrid) method
• Grid search to optimize batch size and number of epochs
• Statistical analysis (t-tests) confirming optimal configurations
• Insights into avoiding overfitting and maximizing accuracy

Hybrid Learning in Neural Networks for Almond Classification

Deriving Backpropagation

This assignment required in-depth derivations of backpropagation update rules for a custom neural network with:

• Unique activation functions in both hidden and output layers
• A custom loss function including a penalty term over non-bias weights
• Specific bias vs. non-bias weight updates

The key objectives included:

• Deriving the update rule for non-bias weights between hidden → output
• Deriving the update rule for non-bias weights between input → hidden
• Comparing bias weight updates vs. non-bias to confirm differences

Backpropagation

Data Augmentation

This study explores data augmentation techniques (e.g., back-translation, random deletion, synonym replacement) to improve machine translation (MT) performance for the low-resource Xitsonga–English language pair. Through cosine similarity scoring and threshold-based filtering, the augmented data retained high semantic fidelity, thereby boosting translation accuracy.

Key Highlights Include:

• EDA & Clustering: Detailed analysis of sentence lengths, word counts, and unsupervised cluster separations (K-Means, t-SNE)
• Data Augmentation: Techniques like back-translation, random word insertion, and thresholding based on cosine similarity ≥ 0.83
• Transformer Models: zaBERT, XLM-R, and RoBERTa showed superior performance over traditional classifiers when trained on augmented data
• Practical Takeaways: Demonstrates the effectiveness of hybrid approaches and semantic filtering in low-resource MT tasks

View Paper (PDF) .

Implementing Electronic Health Records (EHR) in South African Public Hospitals

This academic work proposes an Electronic Health Records (EHR) system to modernize patient-care processes in South African public hospitals. By digitizing patient records, enhancing interoperability, and integrating advanced security measures (e.g., encryption, role-based access control), EHR technology aims to reduce medical errors, improve operational efficiency, and ensure compliance with regulations like POPIA.

Key Highlights Include:

• Automated Patient Registration for real-time data capture and reduced wait times.
• Interoperability & Data Sharing across departments and facilities for integrated care.
• Data Security & Privacy through encryption, audit trails, and multi-factor authentication.
• Clinical Decision Support tools to minimize errors and align treatment with the latest guidelines.
• Action Plan covering phased implementation, staff training, and legal compliance strategies.

View Detailed Paper (PDF)