The Last Mile in AI Deployment

With all the hype around AI, it’s easy to get caught up in the race to build the most sophisticated models, push the limits of computational power, and achieve top scores on renowned research benchmarks. But in these tumultuous times, where AI’s role in society is being hotly debated, scrutinized, and sometimes misunderstood, it’s the last mile—the part about being thoughtful in how we deploy and use AI—that truly matters.

At OpenHouse Research, our recent work with new clients has reinforced the importance of last-mile deployment. The challenges we encountered while onboarding a few new customers due to the removal of Google Universal Analytics data and GA4’s data retention policies have underscored the need for AI solutions that are not just powerful but also practical and context-aware.

In a recent post, we reminded homebuilders of the critical need to back up their valuable GA4 data before it’s automatically deleted after 14 months. This is a perfect example of the importance of overcoming real-world constraints when deploying AI technology—ensuring that data, the key ingredient for context-aware AI, is preserved and ready to be leveraged for future insights.

The Importance of Context in AI Deployment

The impact of AI is profoundly shaped by the context in which it is deployed. In the home-building industry, for example, we often find that while builders may be 80% similar in their operations, the remaining 20% makes each unique.

This 20% difference also gives them a competitive edge in their specific markets. Therefore, a powerful AI system must be deeply rooted in understanding these contextual nuances. Identifying the causal relationships between market demand, local constraints, and a builder’s actions is critical for a model that works well in one market to succeed in another.

Navigating Non-Linear Relationships

However, the causal relationships are often non-linear and complex, making them difficult to decode with basic data science techniques, which might only identify “predictive causality” rather than true causality. Predictive causality can suggest correlations that appear helpful in the short term but might lead to misleading conclusions and ineffective strategies over time.

This is where market-proven approaches from quantitative financial modelling and well-established methodologies from thermodynamics come into play. Dynamical systems theory and differential geometry provide a robust framework for understanding and identifying causal relationships in time series prediction.

Here’s how each concept contributes to this process:

1. State Space Representations Reconstruction: We can infer causal relationships by reconstructing the state space representations of different time series and analyzing how one manifold interacts with the other.
2. Topological Analysis: Topological methods explore the structure of these manifolds. We can identify potential causal relationships by looking for invariant features or changes in the topology.
3. Differential Geometry: Differential geometric techniques allow us to study the curvature and other properties of manifolds. They help us determine how the shape and structure of one manifold are influenced by another, providing further evidence of causality.