Projects

Selected work

6 representative projects across energy forecasting, banking risk, NLP, and policy analytics. All built to production or research publication standards.

SPOTLIGHT

Wattzo Energy Forecasting System

Machine learning for solar & grid demand · Feb–Jul 2026

Problem: Independent solar operators need 12-48 hour forecasts of PV output and grid demand to balance supply safely. Uncertainty quantification critical for confidence in routing decisions.

Solution: Dual-regime LSTM ensemble + conformal prediction. Separate models for clear-sky and cloud-covered periods. SHAP-based feature attribution for operator interpretation.

Results: 18% error reduction vs. baseline; reduced forecast uncertainty bands by 12%. Deployed across 12 sites in Southern Africa.

Tech: PyTorch · TensorFlow · Conformal prediction · SHAP · Production inference API

SPOTLIGHT

Afriland Credit Risk Scorecard

Enterprise risk modeling for retail lending · Mar–Jul 2025

Problem: Afriland First Bank needed systematic credit risk scoring for 15-branch retail loan portfolio (50K+ outstanding loans). Manual decision-making inconsistent and slow.

Solution: Logistic regression scorecard with elastic net feature selection. Threshold optimization for precision/recall tradeoff. 40+ engineered features from transaction history, applicant demographics, repayment patterns.

Results: 82% precision on high-risk flagging. Backtested model identified 80% of future defaults 6 months ahead. Reduced approval-time 4× (40 min → 10 min per application).

Tech: scikit-learn · Excel VBA dashboard · SQL · Shap explainability

SPOTLIGHT

SME Survival Policy Dashboard

Econometric + ML early-warning system for enterprise failure · Apr–Jun 2025

Problem: Cameroon's Ministry of Small & Medium Enterprises needed to identify high-failure-risk firms to target support programs. No systematic data on business survival patterns.

Solution: Hybrid econometric-ML pipeline: Lasso feature selection + ensemble models (CatBoost, Random Survival Forest). Integrated firm-level survival curves with risk scores. Interactive Shiny dashboard for policy makers.

Results: AUC = 0.884 on hold-out test set. Model identified 80% of firms failing 12–24 months ahead. Dashboard adopted by Ministry; now used to guide enterprise credit and training grants.

Tech: CatBoost · scikit-survival · Shiny (R) · PostgreSQL · survival analysis

Also shipped

NATURAL LANGUAGE PROCESSING

SemEval-2026: Polarization in Low-Resource Languages

Multilingual transformers + adaptive focal loss · Published in ACL Proceedings

End-to-end NLP pipeline for identifying polarization manifestations in Hausa and English social media. Ranked 5th globally on Hausa track (50+ teams). Addresses class imbalance via label-wise threshold optimization.

RECOMMENDER SYSTEMS

E-Commerce Product Recommendation Engine

Collaborative filtering + content-based hybrid · 32M interactions

Production recommender trained on 32M user-item interactions. Dual-mode retrieval: ANN indexing for speed, learned-to-rank for relevance. A/B testing shows 12% CTR uplift vs. baseline.

DATA JOURNALISM

Automated Economic Reporter

NLG pipeline for statistical narrative generation

System to automatically generate plain-language economic summaries from statistical tables and forecasts. Integrates template-based NLG with learned headline generation. Used for daily policy briefings.

RESEARCH INFRASTRUCTURE

Master's Thesis: Explainable Forecasting

Conformal prediction + SHAP attribution for uncertainty · Deployed for energy systems

Probabilistic forecasting of electricity consumption and solar PV production with explainable uncertainty quantification. Thesis published as research paper at Deep Learning Indaba 2026.