How I Became an AI/ML Engineer in Nigeria My Real Roadmap Failures and What Actually Worked

The journey was never linear. There were nights I stared at failing training runs wondering if I was wasting my time. Moments I wanted to quit and go back to regular web development. But something kept me going: the belief that we can build AI that actually understands Nigerians in our own voice and context.

Today I serve as Machine Learning and AI Engineer at a fast-growing Nigerian social super-app. This is the detailed, technical story of how it happened.

My Background

I am Emmanuel Ariyo, known online as Ememzyvisuals and founder of Axiveri. For years I focused on web development, building applications such as NaijaPrep for exam preparation, StoreJet for small businesses, automation bots, and more.

I enjoyed shipping web products, but I craved deeper impact. I wanted to create AI systems that could understand and generate Nigerian languages naturally, especially Pidgin with its fluid code-switching.

The Learning Phase Courses vs Real Systems

I started with Udemy courses on machine learning, deep learning, and PyTorch. They gave me solid foundations in data handling, model architecture, training loops, and evaluation metrics.

But courses alone were not enough. The real growth happened when I started building production-grade systems and learning from failures in the wild.

I spent countless evenings and weekends experimenting with fine-tuning open models, curating datasets, implementing robust pipelines, and rigorously evaluating performance on Nigerian speech data.

Key Releases from Axiveri

At Axiveri we have released several open models and datasets that are making a difference:

• NaijaVox-2.0: One of the strongest open multilingual ASR models for Nigerian languages. Built on Whisper-large-v3 with advanced LoRA fine-tuning, SpecAugment, and realistic noise augmentation. Pidgin reaches approximately 85% accuracy with strong code-switching capabilities.

Link: https://huggingface.co/Axiveri/NaijaVox-2.0

• Africlaude-7B: A 7 billion parameter language model specifically tuned for African developers and fintech use cases. It understands Nigerian Pidgin, USSD flows, payment integrations, and CBN compliance contexts natively. It topped our custom AfriclaudeBench evaluation.

Link: https://huggingface.co/Axiveri/Africlaude-7B

• Wazobia-TTS-CC: Our open dataset and related TTS work aimed at high-quality, culturally accurate text-to-speech for major Nigerian languages. This is part of building full speech-to-speech pipelines.

All of these are open under permissive licenses so the community can build upon them.

Landing the Role

The public work and consistent sharing on X and Hugging Face led to the opportunity. After aligning on vision with the Co-Founder and CTO, I joined as independent ML/AI Engineer.

The mandate is exciting: design and implement real-time Pidgin speech-to-speech AI with authentic Nigerian voices for a platform serving over 20,000 users. Responsibilities include user voice data pipelines, model training and fine-tuning, evaluation frameworks, and production API integration.

First payment came quickly, contract signed, and we are already defining the working framework and milestones. Six-month initial term with clear deliverables.

My Practical Roadmap What Actually Moved the Needle

Here is the sequence that worked for me:

1. Master Fundamentals Deeply

Use structured courses (Udemy etc.) to learn concepts fast, then immediately apply them.

2. Identify Painful Real Problems

Focus on gaps in your own environment. For me it was language AI that failed Nigerians in daily use.

3. Ship Small and Iterate Publicly

Release v1 even if imperfect. Learn from community feedback and improve in v2. Public iteration accelerates growth.

4. Emphasize Evaluation and Robustness

Test models under real conditions — noisy environments, code-switching, low-resource settings. A lab model that fails in the field is not useful.

5. Document Everything and Share

Write detailed model cards, post progress, upload datasets and code. Visibility creates opportunities and feedback loops.

6. Stay Consistent Through Doubt

Progress is rarely linear. Keep small daily wins alive even when motivation is low.

Hard Truths and Failures

I wasted time on approaches that did not scale. Early models overfit badly. Compute costs added up quickly on limited budgets. There were long stretches where I felt completely stuck and questioned the entire path.

The turning point was treating every failure as valuable data. Each bad run taught me something about data quality, hyperparameter choices, or architecture decisions.

Advice for Aspiring AI/ML Engineers in Nigeria and Africa

• Start today with whatever hardware and data you have. Kaggle, Colab, and free tiers are sufficient to begin.
• Prioritise problems that matter locally. Global models rarely work perfectly here without targeted adaptation.
• Build in public from day one. Your open work becomes your strongest portfolio.
• Focus on value and community. Share useful resources and genuine connections follow.
• Protect your energy. The journey is long. Celebrate small milestones.

I am still learning every single day. This new role is another chapter in that ongoing education.

Future posts will dive deeper into technical details: dataset curation strategies, training tricks that worked, evaluation frameworks, and lessons from production deployment.

For now you can explore all my projects, models, and earlier writing at https://ememzyvisuals.vercel.app

Let us continue building African AI infrastructure that truly serves us.

Emmanuel Ariyo

Founder at Axiveri | Building NaijaVox, Africlaude, and more