Isaiah Kolawole

Focus

Building data pipelines, forecasting models, and AI applications that turn data into useful decisions, with a focus on reproducibility, data quality, and evidence that people can inspect.

Featured project

Health Programme Analytics — a reproducible Python pipeline for a synthetic facility network, combining reporting validation, a Streamlit dashboard, forecasting comparisons, and evidence-checked AI draft briefs.

• Distinguishes accepted counts, missing reports, and invalid submissions across 216 facility-months, retaining an audit trail.
• Compares three forecasting methods using a time-based holdout. Holt-Winters recorded MAE 27.38 doses and WAPE 3.61% across nine complete synthetic months.
• Supports a PostgreSQL schema and tested SQL views, alongside CSV outputs used by the dashboard.
• Uses local Ollama generation with citation and numeric checks, a deterministic fallback, and human review.

Synthetic demonstration, not real-world programme impact. The same holdout selected and evaluated the model; independent validation remains necessary.

GitHub Code Chat

Local AI codebase assistant — a Python and Streamlit application with a bundled offline example, BM25 keyword retrieval, local embeddings, hybrid rank fusion, and Ollama-generated explanations.

• Reads public repositories at a pinned commit and preserves file/line references.
• Exposes the source excerpts behind answers, with citation-number checks and graceful model-failure handling.
• Includes 20 authored evaluation questions, recorded method comparisons, documented model errors, and automated tests.

Local engineering prototype. Its five-file example evaluation does not establish performance on real repositories or answer accuracy.

InspectAI

Computer-vision inspection workspace — frozen pretrained CNN features, a normal-patch memory bank, ONNX CPU inference, and a FastAPI endpoint with a visual anomaly overlay.

• Fits on 167 normal bottle images; 42 held-out normal images calibrate the review threshold.
• Evaluates both a pixel baseline and CNN method on all 83 official bottle test images.
• Records CNN image AUROC 0.9976, with 2/63 missed defects and 0/20 normal images flagged; includes per-image evidence and both missed contamination cases.
• Processes uploaded images in memory and supports browser inspection history and human review.

Noncommercial research prototype using MVTec AD under CC BY-NC-SA 4.0. Scores are not probabilities; benchmark results do not establish performance on arbitrary camera images.

Tools demonstrated across the projects

Python · pandas · NumPy · scikit-learn · statsmodels · SQL / PostgreSQL · Streamlit · Ollama · RAG / embeddings · PyTorch · ONNX Runtime · FastAPI · Git / GitHub

Project use describes the evidence available in this portfolio; it is not a proficiency rating or professional experience claim.

Explore the work

Interactive demo: isaiah-kolawole-portfolio.vercel.app/demo.html
Source: github.com/kolawoleisaiah325/health-programme-analytics