Purpose

SilicaGuard is an AI-powered public-health platform designed to improve early silicosis risk screening, health education, referral, and disease surveillance among artisanal and small-scale mining communities in Zimbabwe. The project was a Top 20 Finalist at Cimas Healthathon 3.0.

The Problem

Many miners work in high-dust environments without continuous screening, early diagnosis, reliable occupational-health records, or easy access to specialist services. SilicaGuard connects miners, village health workers, hospitals, insurers, and public-health stakeholders through USSD-based self-screening on basic mobile phones, multilingual support in English, ChiShona, and IsiNdebele, AI-powered WhatsApp guidance, a hospital referral-tracking dashboard, district-level risk mapping, and computer-vision-assisted X-ray analysis. It remains an ongoing platform and is not yet nationally deployed.

Tech Stack

Flutter, React, Python, Convolutional Neural Networks, Firebase, Git

Purpose

An explainable machine learning and full-stack credit-risk platform developed using real Zimbabwean loan data, combining a LightGBM and CatBoost ensemble with SHAP explainability to achieve a ROC-AUC of 0.6768 on the Zindi competition leaderboard, predicting loan-default risk and visualising provincial risk patterns for financial institutions.

The Problem

Financial institutions need transparent, accurate loan-default risk assessment. This platform engineers 56 predictive features from borrower, loan, employment, financial, geographic, and repayment data, explains every prediction, and provides applicant-level risk scoring, provincial risk mapping across all ten provinces of Zimbabwe, batch CSV scoring, PDF credit-risk report generation, what-if simulations, role-based access control, offline support, and a multilingual AI assistant supporting English, ChiShona, and IsiNdebele, delivered as a Progressive Web Application.

Tech Stack

Python, LightGBM, CatBoost, SHAP, Pandas, Scikit-learn, FastAPI, Next.js, TypeScript, PostgreSQL, Progressive Web App technologies, Git

Purpose

Kazizi is a smart vehicle diagnostics and predictive-maintenance platform that uses OBD-II and CAN bus data to identify faults, monitor vehicle health, and help drivers detect potential problems before failure occurs.

The Problem

Vehicle owners often discover mechanical problems only after a breakdown. Kazizi collects OBD-II and CAN bus data, interprets fault codes, performs edge processing on a Raspberry Pi, and delivers driver-friendly diagnostic explanations and historical vehicle-performance tracking through a mobile application.

Tech Stack

Rust, Flutter, Raspberry Pi, OBD-II, CAN Bus, SQLite, Embedded Systems, Git

Purpose

A web-based clinical decision-support platform that analyses patient medication reviews to identify sentiment, detect possible adverse drug reactions, generate clinical risk alerts, and support personalised medication recommendations. Developed using a Design Science Research framework and evaluated using the UCI Drug Review Dataset.

The Problem

Patient medication feedback is rarely used systematically to flag risk. This system classifies drug reviews into positive, neutral, and negative sentiment using a TF-IDF and Logistic Regression baseline that achieved approximately 81% classification accuracy, explores BioBERT for healthcare-focused language modelling, and maps sentiment and prediction confidence to clinical risk categories including Mild Concern, Moderate Risk, and Severe Adverse Reaction for both clinician-facing and patient-facing workflows.

Tech Stack

Python, Natural Language Processing, BioBERT, Logistic Regression, TF-IDF, Scikit-learn, FastAPI, Next.js, PostgreSQL, Docker, Git

Purpose

A biometric authentication system that combines face recognition with liveness detection to reduce printed-photo, screen, and video-replay spoofing attacks, using a FastAPI backend, a Next.js frontend, and database storage for registered users and facial data.

The Problem

Face-based access systems are vulnerable to spoofing using printed photos, screens, and video replay. This prototype adds liveness detection alongside face verification, covering face registration, face verification, printed-photo attack detection, and screen-replay attack detection, so that access decisions depend on both identity verification and liveness.

Tech Stack

Python, Computer Vision, Convolutional Neural Networks, OpenCV, FastAPI, Next.js, Git