production · 2026
A retrieval system over the ETO AGORA corpus of global AI governance documents. Every question is split into sub-queries, retrieved in parallel from Pinecone and answered by Gemini 2.5 Flash with sources attached.
production · 2026
A medical RAG chatbot for underserved communities in Northern Nigeria, built for the DataFest Africa 2025 Hackathon. It flags emergencies, takes Hausa voice input and answers from 648 verified health documents.
production · 2026
An architecture advisor that refuses to recommend anything until an interview has gathered real constraints. It then answers from a curated corpus of technical books, with citations back to the source.
production · 2025
An end-to-end model that predicts whether a loan applicant will complete e-signing, a real bottleneck in digital lending. Six models were compared and the winner exported as a .pkl file.
production · 2025
Conversational inventory tracking for building-materials stores. Sales typed in plain language are parsed by Gemini with a regex fallback, saved to Supabase and exposed to Power BI.
production · 2025
Live AI confidence coaching with explainable scoring. A keyword check and a Gemini check both score the user, and the more cautious score wins.
research · 2025
Fraud detection on 1M+ interbank transactions at a 0.3% fraud rate, with no synthetic balancing. A self-audit caught a data-leakage bug and lifted AUC-ROC from 0.82 to 0.92, and SHAP plus Gemini explain every flag.
research · 2025
A convolutional network that classifies 25 American Sign Language hand gestures. Dropout, augmentation and early stopping took test accuracy from 98.65% to 99.78%, with a real-time Flask webcam demo.
research · 2025
A TF-IDF and Naive Bayes baseline for classifying gaming and brand tweets as positive or negative. It reaches 62% accuracy on a small test set, reported as a first baseline with the missing steps named.
research · 2025
Train a model on video-game scripts and lore and it starts writing like one. TF-IDF and Naive Bayes finish a sentence like "the hero must" in the right narrative tone.
research · 2024
What customers buy together and why. Market basket analysis finds the patterns, then item-based collaborative filtering recommends products to the Bronze customer segment.
research · 2024
Cleaning and correlation analysis of a bank campaign dataset to find what tracks with success. The linearity check failed, so the project stopped at correlation insights.
research · 2024
Exploratory factor analysis on perfectionism survey data, then a regression of the latent factors on self-doubt. Run in jamovi, as analysis rather than software.
research · 2024
An ETL pipeline that pulls powerlifting competition data into a normalized PostgreSQL schema, trains a scoped regression (R² 0.411) and feeds Power BI dashboards.
research · 2024
SQL analysis of a bike-sharing system: ridership joined to cost data to compute real profit and find the best time slots, with a Power BI dashboard for operations teams.