Available for AI Engineer roles

I build AI that works on languages
the world forgot.

I'm a BI Architect moving into AI engineering. I build production speech-to-text, translation and local-LLM systems — and I measure them against pre-registered gates before I call anything finished.

See the work GitHub
92.6% → 32.0%
Whisper's word-error rate on Yoruba, then the model that beat it
22
Bible translations shipped in an offline-first reader
$400
Hard compute ceiling for a production AI launch

01 — Flagship

Closing the AI gap for Yoruba

State-of-the-art speech recognition scores a 92.6% word-error rate on Yoruba — essentially unusable. This project fixes that, on a $400 budget, with every number documented in public.

Aso Oke

In development

Yoruba speech-to-text & translation · Solo, built in public

I measured the baseline first, then built against it: a custom Yoruba diacritic normalizer validated on 200 real corpus sentences, a licensed data pipeline (FLEURS, MAFAND, Tatoeba, FLORES-200), and a pre-registered three-path ship gate written before any result existed.

The most interesting outcome is the one that didn't go to plan. An 8-epoch LoRA fine-tune of Whisper large-v3 reached 61.4% WER and failed the pre-registered gate. Meta's Omnilingual ASR LLM 1B, used with zero Yoruba-specific training, scored 32.0% — and shipped. I chose the baseline over my own model, and documented exactly why.

  • 32.0% WER on held-out FLEURS test (817 clips) vs a 92.6% baseline
  • 27.7% WER on own-speaker holdout — evidence the score isn't inflated by data leakage
  • 4.12s p95 latency for one clip on a single T4
  • 60.6 points better than the un-fine-tuned baseline
View the repo Read the build log Ships 13 Dec 2026

02 — AI & Machine Learning

Models, retrieval and evaluation

Systems that do real inference work: local models, retrieval pipelines, document understanding, and question banks generated at scale.

03 — Engineering

Shipped products

Deployed, tested applications — from offline-first React PWAs to CI/CD-pipelineed web apps with validated data.

04 — Try them

Live demos

Everything below is running right now. Click through.

05 — About

How I work

I came to AI from business intelligence — a decade of turning messy data into decisions, and a habit of asking what a number actually means before anyone acts on it. That habit is why this portfolio is built around measurement rather than demos.

My operating principles are simple. Define the bar before you build. Write down the threshold that makes a thing shippable, then let the result decide. Licence before you baseline. Half my architecture decisions on Aso Oke came from licence audits, and one of them (NLLB-200 → MADLAD-400) would have permanently blocked charging for the product if I'd found it at launch instead of in Week 1. Say what doesn't work. The model cards and field reports are as explicit about the gaps as the wins.

I'm currently deep in low-resource speech for Yoruba, and building local-LLM tooling that runs entirely on-device — because privacy, cost and offline capability shouldn't be afterthoughts.

Toolkit

  • Python
  • PyTorch
  • PEFT / LoRA
  • Hugging Face
  • Transformers
  • Gradio
  • FastAPI
  • Flask
  • React
  • TypeScript
  • Tailwind
  • Cloudflare Workers
  • GitHub Actions
  • R / Shiny
  • Databricks
  • W&B