Work
Case studies show how each system was tested and where it falls short. Every figure comes with the data it was measured on, or a Private work label.
6 items, tagged Evaluation
Case studies
Private work
LegalAI Kenya, legal research for advocates
A multi-tenant research workspace over 61,683 passages of Kenyan law that refuses to answer when the evidence is weak.
- AI engineering
- Retrieval
- Evaluation
- Security
- Legaltech
0.303hit@10 on the live index300 litigation questions; settings were tuned on themEvidence: Private work
Public code
Real-time credit-card fraud detection
A LightGBM fraud model, evaluated on transactions it never saw in training, with 1.4 ms p95 in-process model inference.
- ML
- Evaluation
- Fintech
- MLOps
0.80F1 on the time-ordered holdout42,722 held-out transactions; the 0.85 target was missedEvidence for F1 on the time-ordered holdout: Holdout metrics and latency (reports/end_to_end_optimized_results.json) (opens in new tab)
Public code
GEO-AI cropland mapping from satellite data
Crop versus non-crop from Sentinel-1 and Sentinel-2 data in two regions, with an honest look at where the evaluation flattered the model.
- ML
- Data
- Evaluation
- Agritech
0.745held-out accuracy, RandomForest200 points; majority-class baseline 0.703; random splitEvidence for held-out accuracy, RandomForest: Notebook with every run and its output (index.ipynb) (opens in new tab)
Other projects
Customer Churn ML System
Reproducible churn modelling with validation-based model and threshold selection, MLflow tracking, FastAPI serving and drift checks.
- ML
- MLOps
- Evaluation
Structured LLM Sentiment Analysis with DSPy
Structured sentiment analysis and a multi-turn dialogue agent built with DSPy, with experiments tracked in MLflow.
- LLM
- Evaluation
MovieLens Recommender System
Five recommendation approaches compared with ranking and rating metrics, including a hybrid for cold-start users (Moringa School project).
- ML
- Evaluation