Solving retail’s cold-start problem: Malachyte's recommendation reinvention

Malachyte, founded by Sidd, applies advanced AI models—including large language models—to personalize retail recommendations for unknown users, a challenge known as the cold-start problem. The company uses Google Cloud’s Bigtable and Managed Service for Apache Kafka to process real-time user interactions, enabling faster and more accurate predictions. Retailers using Malachyte’s platform have reported sales increases, with some doubling or tripling revenue. The system leverages attention mechanisms similar to those in LLMs to analyze sequences of user behavior, refining predictions with each interaction.

The platform creates a dynamic user vector that updates continuously with each click or query, improving both individual predictions and the broader model. By embedding products and user data into the same vector space, Malachyte’s system aligns user preferences with relevant items in milliseconds. This approach transforms raw behavioral data into actionable insights, enhancing search results and recommendation carousels across product pages and category sections.

To achieve sub-100-millisecond response times, Malachyte relies on Google Cloud’s real-time AI stack, including Kafka for streaming events and Bigtable for fast user vector lookups. The architecture integrates retailer-specific data via Cloud Pub/Sub, reducing integration complexity. Google Kubernetes Engine (GKE) and Compute Engine (GCE) host the platform’s inference and serving layers, ensuring scalability and reliability for high-traffic retail environments.

The system operates as a data cooperative, where anonymized user data from multiple retailers continuously trains the model, improving recommendations for all participants. Malachyte’s migration to Google Cloud demonstrates that production AI systems require more than GPUs and storage—they need real-time infrastructure, streaming layers, and managed services to deliver immediate, personalized experiences at scale.