My Projects

Groundwork — AI Solution Architecture Advisor

An interview-driven architecture advisor that refuses to synthesize a recommendation until a structured elicitation process has gathered enough about real constraints, then grounds its answer in a curated corpus of real technical books — not the model's own pretrained knowledge — with citations back to specific source material. Built solo in two weeks; deployed on real, publicly reachable infrastructure.

Agora — RAG System for AI Governance Documents

Production-grade retrieval-augmented generation system over the ETO AGORA Corpus of global AI governance documents, regulations, and laws. Uses sub-query decomposition, Gemini 2.5 Flash, and Pinecone for grounded, source-cited answers, with a dedicated evaluation dashboard for measuring retrieval and answer quality.

ConfidenceAI — Live Coaching Platform

Live deployment · Real user sessions

Real-time AI confidence coaching with explainable scoring. Not generic motivational text — a hybrid keyword + Gemini confidence assessment (biased toward the more cautious read) drives structured, Pydantic-validated coaching responses, with retry-and-fallback logic so a flaky API call never leaves a struggling user with silence.

Dr. Amina — AI Healthcare for Northern Nigeria

648 health documents · 80% confidence threshold · Hausa voice input

Built a medical RAG chatbot that serves Northern Nigeria's underserved communities. Powered by Google Gemini and Pinecone (648 verified health documents), it detects emergencies with 80% confidence, supports Hausa voice input, and persists conversations via Supabase. Deployed live on Streamlit for the DataFest Africa 2025 Hackathon.

Fighting Fraud with Machine Learning

1M+ transactions · 0.3% fraud rate · 0.92 AUC-ROC

Built a production-grade fraud detection system on 1M+ interbank transactions with a 0.3% fraud rate — no synthetic balancing tricks. Engineered features, trained interpretable models (Logistic Regression → Random Forest), then ran a self-correcting feature-discovery audit that caught a real data-leakage bug and lifted AUC-ROC from 0.82 to 0.92. SHAP + Gemini make every flagged transaction explainable to analysts.

Insightflow — Building Materials Inventory Tracker

Conversational inventory tracking for building materials stores. Parses single and multi-entry sales in natural language with Gemini plus regex fallback, saves transactions in Supabase, and exposes Power BI-ready SQL views for reporting.

AI-Powered Sentence Completion for Video Game Narratives

What happens when you train an NLP model on video game scripts and lore? It starts writing like one. Built a sentence completion engine using TF-IDF + Naive Bayes with emotion-aware generation — give it "the hero must" and it'll finish the thought in the right narrative tone. Hand-curated the dataset with emotional annotations; a Streamlit interface was scoped but not shipped.

Digital Lending E-Sign Prediction System

Production-ready .pkl export · Real fintech workflow

Built an end-to-end ML system that predicts whether loan applicants will complete electronic signing — a real bottleneck in digital lending. Engineered features like income-to-loan ratios and composite risk scores, compared six models (logistic regression through XGBoost/LightGBM/CatBoost) with tuning and overfitting checks, and exported the winner as a .pkl file ready for production integration.

Sentiment Lab — Twitter Sentiment Classification Baseline

A TF-IDF + Naive Bayes baseline for classifying gaming/brand tweets (Borderlands, Call of Duty, Nvidia, and others) as Positive or Negative. 62% accuracy on a small held-out test set — reported honestly as a first baseline, not a production system, with the incomplete steps (full-dataset training, 4-class handling, actual validation-set evaluation) named explicitly.

Sign Language Hand Gesture Recognition with CNN

99.78% accuracy · 25 ASL gestures

Trained a convolutional neural network to classify 25 American Sign Language hand gestures, improving from 98.65% to 99.78% test accuracy by tackling overfitting with dropout, data augmentation, and early stopping. Built a real-time Flask webcam demo, then started (but didn't finish) a browser-based deployment to fix the local-webcam-only limitation.

Bike-Sharing Rider Efficiency & Profitability Analysis

Used SQL to reverse-engineer a bike-sharing system's operations — joining ridership against cost data to compute real profit, finding the time slots and seasons where efficiency meets profitability, and preparing the lagged features a future demand-forecasting model would need. Built an interactive Power BI dashboard so operations teams can see when and where to deploy bikes instead of guessing.

Market Basket Analysis & Collaborative Filtering

Explored what customers buy together and why — using market basket analysis to find purchasing patterns, then item-based collaborative filtering to generate personalized recommendations for the Bronze customer segment. The goal was turning raw transaction logs into "customers who bought X also bought Y" intelligence that a retail team can act on.

Data Predictive Insights ETL Pipeline

Built a full ETL pipeline that pulls powerlifting competition data, transforms it through Python, and loads it into a normalized PostgreSQL schema for analysis. Trained a scoped linear regression model to predict total weight lifted (R² 0.411), and connected the whole thing to Power BI for interactive dashboards.

Bank Marketing Insights — Data Cleaning & Correlation Analysis

Cleaned and analyzed a bank's marketing campaign dataset (client demographics, campaign specifics, economic indicators) to find which factors actually track with campaign success. Ran a full correlation study, tested whether the data supported Multiple Linear Regression, and made the call to stop at correlation-based insights once the linearity assumption failed rather than force a model that didn't fit.

Factor Analysis & Linear Regression on Perfectionism Data

Applied exploratory factor analysis to a psychology dataset on perfectionism dimensions — reducing a complex web of survey responses into meaningful latent factors. Then used multiple linear regression to model how those hidden factors predict self-doubt. Statistical work done in jamovi, not code — the kind of analysis that turns messy behavioral survey data into clear, actionable patterns.