Jump to: About • Zindi • Focus • Competencies • Tech • Competitions • Projects • GitHub • Connect
- 🔭 Building and sharing ML, Data Science, and Data Engineering work at africdsa.com
- 🏆 Zindi competitor — consistently placing in top rankings across African AI challenges
- 🌱 Currently deepening expertise in NLP, Speech Recognition, and LLM fine-tuning
- 🚀 Building full-stack data products: FastAPI + Next.js + Supabase pipelines
- 💬 Ask me about Python, ML workflows, data pipelines, geospatial ML, or real estate analytics
- 🌍 Proudly coding from Nairobi, Kenya — solving African problems with AI
- ♟️ Chess: daily tactics, endgames, and openings
- Building: Nairobi Property Pricing — full-stack real estate intelligence platform (Next.js + FastAPI + Supabase)
- Competing: Active Zindi challenges — NLP, computer vision, time-series forecasting
- Exploring: LLM fine-tuning and speech recognition for low-resource African languages
- Sharing: ML tutorials and open-source African data science at africdsa.com
I build practical, production-minded ML and data systems — from ingestion and transformation to modeling, evaluation, and deployment.
|
Machine Learning |
Data Science |
Data Engineering |
| Competition | Platform | Position | Solution |
|---|---|---|---|
| Barbados Lands & Surveys Plot Automation | Zindi | 🏅 Top Finish | Repo |
Full-stack real estate intelligence platform for Nairobi — automated data pipeline from scraping to interactive affordability dashboard.
What makes it strong
- Automated scraping pipeline: daily GitHub Actions workflow scrapes live listings and pushes to Supabase
- Intelligent parsing: extracts bedroom counts from messy titles/URLs (
"2br","two bed", etc.) - Affordability analytics: price-per-bedroom metrics, tier segmentation, location summaries
- Interactive frontend: Next.js + TypeScript dashboard with geographic affordability map
🔗 Backend: josephgitau/nairobi_property_pricing 🔗 Frontend: josephgitau/nairobi-property-pricing-frontend 🌍 Live: josephgitau.me
End-to-end geospatial digitization pipeline for cadastral survey maps: parcel boundary detection + polygon extraction + OCR to produce structured, searchable outputs ready for indexing and downstream GIS use.
What makes it strong
- Boundary segmentation: UNet++ with EfficientNet-B7 encoder for precise parcel detection from raster maps
- Production-grade post-processing: polygon cleaning (hole removal, simplification, smoothing) for valid GIS geometries
- Robust OCR strategy: Qwen3-VL-30B (vision-language) for noisy map text — zero/few-shot generalization instead of risky fine-tuning
- Reproducible deliverables: training + inference notebooks, checkpoints, and final merged outputs (polygons + OCR text)
Results: Public 0.965006861 • Private 0.970242006
🔗 Repo: josephgitau/Barbados-Lands-and-Surveys-Plot-Automation-Challenge 📒 Data prep: Open in Colab
NLP-powered bot that generates contextual stories based on detected sentiment.
🔗 Repo: josephgitau/Sentiment_Story_Generation_Bot
- ♟️ Chess is both relaxation and mental training — tactics, endgames, and openings every day.
- 🌍 Passionate about applying AI to solve African problems — housing, agriculture, language.
- 🧩 Love clean, elegant solutions to messy real-world data.
- 🚀 Always building something new — the best model is the next one.
"Build from data. Compete with purpose. Ship it."