Q1. Could you start by giving us a brief overview of your professional background, with a particular focus on your industry expertise?

I've spent over 20 years working with large-scale distributed systems and currently lead teams as an AI systems architect, focusing on building and managing enterprise technology platforms.

While serving as IBM’s Vice President for AI Engineering, I worked on designing and rolling out AI platforms for industries that face strict regulations. My main interests were making sure these systems could scale, training models across different servers, handling cloud integration, and always keeping governance in mind.

Along the way, I've worked on everything from integrating large AI models, optimizing how GPU clusters run, tracking where data comes from, and making AI decisions easier to explain—especially in industries that need to follow strict rules. I try to connect the latest AI research with practical solutions that businesses can actually use.

 

Q2. How is the global energy shortage impacting the roadmap for regional AI data centers in emerging markets like Nigeria?

Unstable energy supplies have a big impact on how we build AI systems. In countries like Nigeria, designing AI infrastructure means thinking about things like:

• Intermittent power availability
• High PUE (Power Usage Effectiveness) in warm climates
• Voltage instability impacting GPU reliability
• Cooling constraints limiting rack density

This shifts architecture toward:

• Smaller, modular GPU units
• Fault-tolerant distributed training with checkpoint resilience
• Energy-aware workload scheduling
• Inference-heavy architectures over large-scale pretraining
• Hybrid renewable + storage-backed power design

The hardest part isn’t just making sure there’s enough electricity—it’s making sure it’s reliable. This affects everything, from how we set up computer clusters to how we prepare for outages and backups.

 

Q3. In highly regulated sectors (Banking/Insurance), what is the 'Auditability Gap'? Can an enterprise truly explain an AI-driven credit denial to a regulator today?

The ‘auditability gap’ shows up when AI models get so complicated that today’s tools can’t really explain their decisions. For example, in deep learning systems that help decide credit risk, we run into challenges like:

• Non-linear feature interactions
• High-dimensional embedding spaces
• Opaque attention mechanisms
• Dynamic retraining pipelines

Tools like SHAP and LIME can show what influenced a decision, but they can’t always explain exactly why it happened.

To close the gap technically, enterprises need:

• Model lineage tracking (training dataset versions, hyperparameters)
• Immutable audit logs
• Feature store traceability
• Model behavior testing under counterfactual simulation
• Governance integrated into CI/CD pipelines

If you want real auditability, you need strong engineering behind the scenes—not just fancy dashboards.

 

Q4. Many firms claim their 'proprietary data' is a moat. From your board-level view, how much of that data is actually 'AI-ready' versus 'digital landfill'?

When we review company data, most of it isn’t ready for AI yet because of:

• Schema drift across systems
• Inconsistent labeling guidelines
• Lack of feature normalization
• Poor metadata tagging
• Data silo fragmentation

AI-ready data must satisfy:

• Consistent schema governance
• Time-series integrity (no leakage)
• Clean joins across business domains
• Clear usage rights for training
• Automated validation pipelines

If you skip these steps, you can end up with data issues like bias or unstable performance. The better your data is prepared, the more reliable your AI and the quicker you can improve it.

 

Q5. In your experience, what percentage of enterprise GPU CapEx is currently poorly optimized due to bottlenecks in data orchestration versus actual compute limits?

A lot of companies don’t get the most out of their GPUs because of:

• I/O bottlenecks (slow object storage reads)
• CPU-GPU imbalance
• Network latency in distributed training
• Poor sharding strategies
• Poorly performing data loaders
• Idle time during preprocessing

Many times, the GPUs just sit idle, waiting for data to catch up.

Optimizing:

• High-throughput storage (NVMe, parallel file systems)
• Data locality strategies
• Mixed-precision training
• Efficient batch sizing
• Cluster orchestration (e.g., Kubernetes-based GPU scheduling) 

Simple changes like these can make systems 20 to 40 percent more efficient, all without buying extra hardware.

Usually, it’s the data pipeline that slows things down—not the actual computing power.

 

Q6. What are some technical reasons a Generative AI pilot fails to reach full-scale production?

Most failures happen at the points where different parts of the system have to talk to each other.

Common technical blockers:

• Weak RAG pipelines (poor embedding selection, weak indexing)
• Hallucination due to insufficient domain grounding
• Latency spikes under concurrency
• Lack of horizontal scaling architecture
• No observability layer for model drift
• Cost blowouts due to token inefficiency

Production-grade generative AI requires:

• Retrieval optimization
• Structured prompt orchestration
• Guardrails and validation layers
• Feedback-driven reinforcement loops
• Full-stack MLOps integration

Most pilot projects don’t work out because people focus too much on the model, and not enough on how the whole system fits together.

 

Q7. If you were an investor looking at companies within the space, what critical question would you pose to their senior management?

If I were looking at an AI company, my first question would be: What’s the biggest thing slowing your system down, and can you actually measure it

Specifically:

• What is your GPU usage rate under production load?
• What percentage of inference latency is retrieval vs. generation?
• What is your retraining cycle time?
• How do you detect and mitigate model drift automatically?
• What happens to performance if input volume doubles overnight?

The companies that can answer these questions with real numbers—not just theories—are the ones that are truly ready to scale up.