Brain Tumour DetectionPaper in review
A custom CNN and a fine-tuned ResNet for classifying tumours in MRI scans, later extended into a federated setup so hospitals can train together without sharing scans. 98.5% accuracy with EfficientNetB0.
EcoFlow AITop 25 of 2000+ · KFUPM
An event-driven digital twin on Google Cloud that reads live IoT data to predict crowd surges and trim carbon output. Firestore for the real-time side, BigQuery for the analytics.
TrustLock2nd · Wema Bank
An access gateway for remote teams. Role-based permissions, JWT auth, and an audit trail that records IP, device and location for every sign-in.
Padipay
A payment system you operate by speaking to it, built to understand Nigerian English as well as standard English. Voice recognition on the front, ordinary transaction validation behind it.
CropWatch
Crop disease detection from photos of leaves, using transfer learning on a CNN trained in two stages — features frozen first, then fine-tuned. The backend serves inference to a mobile app in the field.
GPT-2, rebuilt and fine-tuned
I rewrote GPT-2 in PyTorch to understand it properly, then fine-tuned it on my own dataset for sentiment classification and spent a while getting inference latency down.
HelpBot
A document assistant that reads PDFs and answers questions about them through the OpenAI API. Trimming the prompts and payloads cut response time by 65%. My ALX final project.
Fraud Detection
A model that flags suspicious transactions from spending patterns, wrapped in a small Flask interface for simulating them.
Facial Emotion Recognition
Reads emotion from a face — happy, sad, angry, surprised — with a web app for uploading a photo and seeing what it thinks.
ResearchIn review.
Lightweight and Privacy-Preserving Federated Deep-Learning Framework for Efficient Brain Tumor Detection Across Distributed Medical Centers
Mustapha O., Charles O. · 2025
Hospitals can't pool patient scans, so the model has to travel instead of the data. This framework trains across medical centres without any scan leaving its own building. A custom CNN reached 84% accuracy and EfficientNetB0 reached 98.5%; pruning and quantisation brought it down to something that runs on hardware a clinic already owns.
