Engineering Intelligence: Keynote Talk at ML Lagos Community Day 2025 · GitHub

Great models in bad systems fail; average models in well-engineered systems succeed. The talk drives home that a model is just a component, and intelligence only emerges when the entire ML pipeline works together. It strips away the hype around model accuracy and puts the focus on the engineering that keeps predictions reliable at scale.

The common mistake is to treat model training as the end of the project. In reality training is about ten percent of the effort; the remaining ninety percent is spent building data ingestion, feature generation, inference services, decision logic, and continuous monitoring. Real-world examples like Netflix recommendation, Google Search ranking, and fraud detection systems illustrate that no one relies on perfect models- they rely on robust systems that handle drift, bias, and data bugs.

When the data pipeline feeds buggy data, the model learns those bugs and amplifies bias. Training-serving skew occurs when the code path used at inference differs from training, leading to silent degradation. Observability must go beyond logs and metrics to watch input and output distributions, decision rates, and business impact. Without feedback loops the system never improves and errors compound.

MLOps is framed as systems engineering: ownership of data quality, model performance, and incident response is essential. Human-in-the-loop processes catch edge cases, provide corrections, and build trust. Designing for failure-handling bad inputs, missing features, traffic spikes, and partial outages-prevents the worst-case silent failures where predictions drift without alerts.

The final advice is to build boring, reliable intelligence. Favor simple models with strong defaults and clear fallbacks over complex, fragile pipelines. Stability, maintainability, and observability outweigh marginal accuracy gains. By treating ML as an engineered system, teams can deliver real business value and keep their models alive in production.