Enterprise AI Infrastructure for Africa

The AWS AI Practitioner is the entry point. It covers foundational concepts — what machine learning is, how AI services like Rekognition, Comprehend, and Translate work at a high level, and when to apply them. It is broad rather than deep, designed to give you the vocabulary and a mental model for how AWS structures its AI offering. I found it useful for framing conversations with stakeholders who want to understand what AI can do without getting into the engineering.

The Machine Learning — Specialty goes deep into the technical pipeline: data engineering, exploratory analysis, modelling, and operationalising ML workloads on AWS. It covers SageMaker extensively — training jobs, hyperparameter tuning, inference endpoints, and model monitoring. The curriculum assumes you understand the mathematics behind common algorithms (linear regression, decision trees, clustering) and can reason about which approach fits a given problem. This is the certification that forces you to think like an ML engineer, not just a cloud architect.

The Machine Learning Engineer — Associate sits between the two in terms of audience but is more operationally focused. It emphasises MLOps: building reproducible pipelines, automating model retraining, deploying to production with canary rollouts, and monitoring for drift. Where the Specialty cert asks "can you build a model?", the Engineer cert asks "can you keep it running reliably at scale?" For platform engineers, this is arguably the most directly applicable of the three.

Across the DeepLearning.AI short courses — particularly Knowledge Graphs for RAG, Embedding Models: from Architecture to Implementation, and Prompt Compression and Query Optimization — a clear picture emerges of what a production RAG system actually requires. It is not just "plug a vector database into your LLM." The curriculum walks through embedding model selection, chunking strategies, retrieval ranking, and the trade-offs between dense and sparse retrieval methods.

The Knowledge Graphs for RAG course (built with Neo4j) was particularly eye-opening. It demonstrates how graph-based retrieval can surface relationships between entities that flat vector search misses entirely. For enterprise use cases — compliance documents, internal policies, interconnected product catalogues — this matters. The Prompt Compression course from MongoDB covered a different angle: how to reduce token costs and latency by compressing retrieved context before it reaches the LLM, without losing the information the model needs to generate accurate responses.

The LangChain for LLM Application Development course tied these components together into application patterns — chains, agents, and retrieval pipelines that can be composed and tested. For anyone building RAG systems in an enterprise context, these courses provide a solid, vendor-neutral foundation for understanding the architecture decisions you will face.

TOGAF — The Open Group Architecture Framework — is often associated with traditional enterprise IT: application portfolios, integration patterns, governance boards. Having completed both the TOGAF Enterprise Architecture Practitioner and the Applied TOGAF Practitioner certifications through The Open Group, I've found that the framework is more relevant to AI programs than most technical practitioners expect.

AI initiatives fail more often from organisational misalignment than from technical limitations. TOGAF's Architecture Development Method (ADM) provides a structured way to move from business motivation through to technology implementation — ensuring that an AI platform is built to serve actual business capabilities, not just technical curiosity. The capability mapping exercises in TOGAF are directly applicable: which business functions benefit from ML, what data flows need to exist, what governance is required for model outputs that affect customers.

In the African enterprise context, this matters even more. Many organisations are at an earlier stage of AI adoption, and the temptation is to jump straight to tooling — spinning up GPU clusters, deploying models — without the architectural groundwork. TOGAF provides a common language between technical teams and business leadership that prevents the AI platform from becoming an expensive experiment that never reaches production. Architecture thinking is not overhead; it is the difference between a proof of concept and a production capability.