papers/hanzo-ai-commerce/sections/02-background.tex

\section{Background}


\subsection{E-Commerce AI Evolution}

AI in e-commerce has evolved through three generations:

\textbf{First generation (1995--2005): Rule-based systems.} Simple heuristics drive recommendations (``customers who bought X also bought Y''), fraud rules (velocity checks, blacklists), and pricing (cost-plus margins).

\textbf{Second generation (2005--2015): Specialized ML.} Machine learning models improve specific functions. Amazon's item-to-item collaborative filtering \cite{linden2003} revolutionized recommendations. PayPal deployed neural networks for fraud detection \cite{paypal2013}. These systems operate independently with limited cross-functional learning.

\textbf{Third generation (2015--present): Integrated AI.} Modern systems unify ML infrastructure across functions. Alibaba's AI platform \cite{alibaba2017} demonstrates system-wide intelligence. Hanzo extends this approach to a platform available to any merchant.

\subsection{Machine Learning Infrastructure}

Production ML requires infrastructure beyond model training:

\begin{itemize}
\item \textbf{Feature stores}: Centralized repositories for feature computation and serving.
\item \textbf{Model registries}: Version control for trained models.
\item \textbf{Serving infrastructure}: Low-latency prediction at scale.
\item \textbf{Monitoring}: Drift detection and performance tracking.
\end{itemize}

Hanzo builds upon emerging standards (Feast, MLflow, TensorFlow Serving) while adding commerce-specific extensions.