Final-year Computer Science student at the University of Lagos, researching African NLP and building agentic AI systems.
My work sits between research and applied engineering. On one side, I build language technology for low-resource African languages, mostly Yorùbá, covering diacritic restoration, AI-text detection, and translation-quality evaluation. On the other, I build agentic systems that move past static text processing into reasoning, planning, and autonomous execution. I co-authored a paper at the EACL 2026 AfricaNLP workshop and keep a handful of models and demos on the Hugging Face Hub.
Alongside the building, I write. Putting ideas into words is how I think through problems, and I publish short essays on applied AI and African NLP on Medium.
| Project | What it is | Stack | Links |
|---|---|---|---|
| Dyon | A source-available Python framework (~9,600 LOC) for building agentic cross-domain digital twins, validated by a three-twin Salesman / Customer / Product system. | Python, LangChain, RL, Eclipse Ditto | PyPI · GitHub |
| Yo-ByT5 | A byte-level T5 that restores tone marks and underdots to plain Yorùbá text, trained on Google Cloud TPUs. 83.8% word accuracy, 83.9 ChrF. | ByT5, TensorFlow, TPU v6e-8 | Demo · Model |
| Yorùbá AI-Text Detector | A transformer classifier separating human-written Yorùbá from LLM output, among the first such systems for a low-resource tonal language. | PyTorch, Hugging Face, Gradio | Demo |
| Oracle | An autonomous agent for a live game economy (Torn City) that ingests data streams, holds persistent state, and acts over the Discord API. | Python, Discord API, state management | Private |
| FER2013 Classifier | Benchmarked CNN architectures for facial-expression recognition with ablations on depth and regularisation, then deployed the best model. | PyTorch, fast.ai, CNNs | Demo |
Trust but Check: LLM-Assisted Review of Human Translations in African Languages EACL 2026, AfricaNLP Workshop (published 27 Jan 2026). Co-authored work using multilingual LLMs to audit translation quality across 11 low-resource African languages, combining LLM-as-a-judge scoring with native-speaker review on the MAFAND-MT test set. Read on OpenReview
- Languages: Python, R, Java, C++
- Deep learning: PyTorch, TensorFlow, fast.ai
- NLP & ML: Hugging Face Transformers, ByT5 and seq2seq models, scikit-learn
- Agentic AI: LangGraph, LLM tool-use, state-aware agents
- Data & infrastructure: NumPy, pandas, Google Cloud TPU (v6e-8), Git, SQL
Open to research collaborations, contract work, and full-time roles in NLP and agentic AI engineering.
Email · LinkedIn · X · Medium · Hugging Face