Four flagship projects, each with its own page: why I built it, the problems I hit, the decisions I made and the results. Then everything else I’ve built.
Looks at a breast scan and the patient’s clinical record together, and suggests one of three diagnoses.
// research project, not a clinical tool.
// why I built it · 3 decisions · results
Sorts breast-ultrasound images into benign, malignant or normal — and shows why a deep model was worth it over a classic pipeline.
// research project, not a clinical tool.
// why I built it · 3 decisions · results
Reads Nigerian business news four times a day and turns it into structured ownership records and a daily executive brief.
// why I built it · 4 decisions · results
Replaced manual copying of polling-unit records with a scraper that produces clean, structured workbooks for every state.
// why I built it · 4 decisions · results
IssueFine-tuning all of MedCLIP overfit almost at once. Validation loss climbed after the first epochs — a few hundred patients can’t re-teach a foundation model.
DecisionKept the pretrained encoder frozen and trained only the heads.
ResultStable training and better generalisation.
02 · ultrasoundIssueThe baseline missed too many cancers. It caught only 55% of malignant cases — not acceptable for screening.
DecisionUsed EfficientNet-B0 transfer learning on the same balanced data.
Result98% malignant recall and 95.75% accuracy.
03 · auraIssueIt once published junk while every run showed green. An earlier version silently fell back to keyword heuristics and, for days, published sports and celebrity stories as corporate intelligence.
DecisionFail loudly: if every provider fails, the run exits non-zero and publishes nothing. Every record now stores which model produced it.
ResultNo more quiet bad data; failures are visible.
04 · pollingIssueThe source data was mislabelled. Some “LGA” columns actually contained State names.
DecisionAdded a semantic mapping layer that corrects these during extraction.
ResultA consistent hierarchy instead of hand-fixing every sheet.
A supervised Vision Transformer against self-supervised DINOv2, with Grad-CAM to show where each model looks.
ViT · DINOv2code ↗ // deep learningRecognises human activity from wearable-sensor data (HAR-70) with a Transformer encoder.
PyTorchcode ↗ // llm & ragSemantic search, RAG over ChromaDB, LLM evaluation and tool-calling agents, ending in a full-stack capstone.
LangChain · FastAPIcode ↗ // full-stackPersonal finance dashboard; an LLM workflow logs spending from bank-alert emails. Works offline as an app.
React · Supabasecode ↗ // clinical mlRisk classifier evaluated on ROC-AUC and recall, so risky patients are not missed.
scikit-learncode ↗ // security + mlGamified training against scams with 1,600+ Nigeria-focused scenarios, plus a spam-detection model.
1,600+ scenarioscode ↗ // webMulti-page responsive redesign built during the Kontemporary Konsulting internship.
HTML · CSS · JScode ↗ // webMulti-page client website for a finance firm, with a FastAPI contact backend.
FastAPIcode ↗ // webPortfolio website built for a cybersecurity professional.
HTML · CSS · JScode ↗Three PyTorch studies; ViT and DINOv2 reach 0.96–0.99 test accuracy on 4-class brain MRI.
0.96–0.99code ↗ // medical imagingEfficientNet-B0, ResNet50, SVM, XGBoost, Random Forest and Decision Tree compared on BUSI.
6 modelscode ↗ // medical imagingEfficientNet-B0 and ResNet50 on 8-class endoscopy images, with Grad-CAM.
0.91 accuracycode ↗ // computer visionResNet-18 transfer learning on 11,000 images.
~95% val. accuracycode ↗ // medical imagingDenseNet121 transfer learning on 2,000 ECG images.
DenseNet121code ↗ // time seriesTransformer encoder on windowed accelerometer data.
Transformercode ↗ // fraudRandom Forest with SMOTE on highly imbalanced transactions.
SMOTEcode ↗ // forecastingHybrid SARIMA + LSTM time-series forecasting.
SARIMA + LSTMcode ↗