AI Engineer β 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.
π Portfolio: https://leye-ajibola.netlify.app π« Email: leyeajibola@gmail.com πΌ LinkedIn: https://www.linkedin.com/in/donne4real
This page is written for engineers, so here's the two-minute version in plain terms.
What I actually do. I make software that understands speech and language, and I prove it works by measuring it against a number I set before building it β not a number I picked afterwards to look good.
The flagship example, without the jargon. The most widely used speech-recognition model in the world is effectively unusable for Yoruba: it gets roughly 9 out of every 10 words wrong. I measured that, then found a model that gets it down to about 3 in 10, and I published the evidence β including the parts where my own attempts failed.
Three terms you'll see here:
- WER (word-error rate) β how often the model gets a word wrong. Lower is better.
- Benchmark / baseline β the existing system you're measuring yourself against.
- Ship gate β the pass mark I write down before I start, so the result can't be fudged.
Every number below links to the repository where you can check it.
Closing the AI gap for Yoruba. OpenAI's Whisper large-v3 scores a 92.6% word-error rate on Yoruba β essentially unusable for tens of millions of speakers.
I measured the baseline first, then built against it: a custom diacritic normalizer (a cleaner that turns typed accents like "α»Μ" into plain letters so text can be compared reliably), a per-dataset licence-audited data pipeline, and a pre-registered ship gate (the pass mark written down before the work started).
The honest outcome is the interesting one β and it is two failures, not 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 β failed the gate |
| Omnilingual ASR LLM 1B, as-is | 32.0% | zero Yoruba training β cleared the ship bar |
I chose the baseline over my own model, and documented exactly why. It also held up on unseen local speakers (27.7% WER) and on women's voices (32.8%) β FLEURS' Yoruba test split is 100% male, so that number did not exist until I measured it.
The translator failed both of its own gates and launches with a visible disclaimer rather than as a polished feature: BLEU 11.7 / 8.0 against a pre-registered bar of 25 (BLEU is the standard 0β100 score for translation accuracy), and a p95 of 7.4s against a 3s limit (p95 means 95% of requests finish faster than this). A cheap test on Nov 9 ruled out the obvious suspect (beam search) and pointed at the 4-bit quantisation instead.
4.12s p95 latency on a single T4, a $400 hard compute ceiling ($149.95 spent so far), and a public decision log.
β github.com/donne4real/aso-oke Β· model adapter Β· ships 13 Dec 2026
| Project | What it is | Link |
|---|---|---|
| Bible in African Languages | Bible reader that works with no internet connection, 22 translations, 45 tests | bible-app.leyesapps.workers.dev |
| Bible in African Languages β Google Play | Published Android app, 50+ downloads. No data collected, nothing shared | com.wordupafrica.biblereader |
| NGX Tracker | ~146 Nigerian Exchange stocks, tracked live | ngx-tracker.pages.dev |
| Cert Mastery | 300 questions each for AWS, Security+, CISSP, PMP, NCE, NCMHCE β the first three are the ones I'm certifying on | cert-mastery.netlify.app |
| Powerball Predictor | Looks through 296 past lottery draws for patterns β hot numbers, 'due' numbers and pairs | powerball-predictor.netlify.app |
| Exam Prep NG | WAEC / NECO / JAMB practice questions across 12 subjects, checked automatically before every publish | donne4real.github.io/exam-prep-ng |
| PocketOffice | Writer, Calc, Impress, text editor and PDF tools in a browser | donne4real.github.io/PocketOffice |
| Vocabulary App | Vocabulary practice on a schedule that makes words stick | build-my-vocab.netlify.app |
| Nia The Investigator | Interactive investigation experience | niatheinvestigator.netlify.app |
| Project | What it is |
|---|---|
| Aso Oke | Speech-to-text and translation for Yoruba, custom model tuning, pass marks set in advance |
| DataInsight | Offline local-LLM data analyst β 100% on-device, no API key |
| Exam Prep NG | Scripted curriculum pipeline generating question banks at scale |
| MarketSage | Stock screener across 10 investment philosophies |
Bible in African Languages Β· Cert Mastery Β· NGX Tracker Β· PocketOffice Β· Powerball Predictor Β· Vocabulary App
AI/ML β Python Β· PyTorch Β· PEFT/LoRA Β· Hugging Face Β· Transformers Β· Gradio Β· FastAPI Β· Flask Β· W&B Frontend β React Β· TypeScript Β· Tailwind Β· Cloudflare Workers Β· JavaScript Β· PWA Data/Infra β GitHub Actions Β· Cloudflare Pages Β· Netlify Β· R/Shiny Β· Databricks Β· SQL
- Define the bar before you build. Write the pass mark down first, then let the result decide.
- Check the licence before you baseline. Half my architecture calls on Aso Oke came from licence audits β 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. Automated checks that validate data before every build (CI/CD); 45-test suites; accessible, responsive UIs.
- π¨ Building β Aso Oke, shipping 13 Dec 2026
- π Certifying β CISSP, Security+ and AWS, over the next year
- π Learning β AWS AI & ML Scholars Nanodegree
- π Based in β Virginia, USA, working remotely
Every number on this page is measured, not estimated. Source for all claims: the repositories above.