AI Engineer — production speech-to-text, translation, and local-LLM systems.
I'm a BI Architect moving into AI engineering. I build systems that do real inference work, and I measure them against gates defined before I start building.
🌐 Portfolio: https://leye-ajibola.netlify.app 📫 Email: leyeajibola@gmail.com 💼 LinkedIn: https://www.linkedin.com/in/donne4real
Closing the AI gap for Yoruba. OpenAI's Whisper large-v3 scores a 92.6% word-error rate on Yoruba — essentially unusable for tens of millions of speakers.
I measured the baseline first, then built against it: a custom diacritic normalizer, a per-dataset licence-audited data pipeline, and a pre-registered three-path ship gate.
The honest outcome is the interesting one:
| Model | FLEURS Yoruba test WER | Note |
|---|---|---|
| Whisper large-v3, no fine-tune | 92.6% | the floor |
| Whisper large-v3 + LoRA, 8 epochs | 61.4% | my own fine-tune — failed the gate |
| Omnilingual ASR LLM 1B, as-is | 32.0% | zero Yoruba training — shipped |
I chose the baseline over my own model, and documented exactly why. 4.12s p95 latency on a single T4, a $400 hard compute ceiling, and a public decision log.
→ github.com/donne4real/aso-oke · ships 13 Dec 2026
| Project | What it is | Link |
|---|---|---|
| Word Up | Offline-first Bible reader, 22 translations, 45 tests | bible-app.leyesapps.workers.dev |
| NGX Tracker | ~146 Nigerian Exchange stocks, tracked live | ngx-tracker.pages.dev |
| Cert Mastery | 300 questions each for AWS, Security+, CISSP, PMP, NCE, NCMHCE | cert-mastery.netlify.app |
| Powerball Predictor | Hot/cold + due-number analysis over 296 draws, 4 weighted strategies | powerball-predictor.netlify.app |
| Exam Prep NG | WAEC / NECO / JAMB question banks, 12 subjects, CI/CD-validated | donne4real.github.io/exam-prep-ng |
| PocketOffice | Writer, Calc, Impress, text editor and PDF tools in a browser | donne4real.github.io/PocketOffice |
| Tappers' Apprentice | Interactive Nigerian-set story, custom JS engine | donne4real.github.io/tappers-apprentice |
| Daily Plan | Scheduling and task views | leyes-daily-plan.netlify.app |
| Vocabulary App | Spaced-repetition vocabulary building | build-my-vocab.netlify.app |
| Nia The Investigator | Interactive investigation experience | niatheinvestigator.netlify.app |
| Project | What it is |
|---|---|
| Aso Oke | Yoruba STT + translation, LoRA fine-tuning, pre-registered gates |
| DataInsight | Offline local-LLM data analyst — 100% on-device, no API key |
| DocSummarization | PDF and multi-file summarisation with a Streamlit UI |
| Exam Prep NG | Scripted curriculum pipeline generating question banks at scale |
| MarketSage | Stock screener across 10 investment philosophies |
Word Up · Cert Mastery · NGX Tracker · PocketOffice · Powerball Predictor · Vocabulary App
AI/ML — Python · PyTorch · PEFT/LoRA · Hugging Face · Transformers · Gradio · FastAPI · Flask · W&B Frontend — React · TypeScript · Tailwind · Cloudflare Workers · JavaScript · PWA Data/Infra — GitHub Actions · Cloudflare Pages · Netlify · R/Shiny · Databricks · SQL
- Define the bar before you build. Write the ship threshold down first, then let the result decide.
- Licence before you baseline. Half my architecture calls on Aso Oke came from licence audits — one of them would have permanently blocked monetising the product if I'd found it at launch instead of in Week 1.
- Say what doesn't work. The model cards state the gaps as plainly as the wins.
- Ship the boring parts too. CI/CD that validates data before every build; 45-test suites; accessible, responsive UIs.
- 🔨 Building — Aso Oke, shipping 13 Dec 2026
- 📚 Learning — AWS AI & ML Scholars Nanodegree
- 🌍 Based in — Nigeria, working remotely
Every number on this page is measured, not estimated. Source for all claims: the repositories above.