Malachyte, an AI-powered ecommerce recommendation platform founded by Sidd, has developed a solution to retail's "cold-start" problem—the challenge of recommending products to first-time or little-known users. The platform applies neural networks with attention mechanisms (the same concept powering large language models) to personalize retail search and product pages by analyzing the sequence of customer interactions rather than relying solely on user profiles or historical data. Using Google Cloud infrastructure including Bigtable and Managed Service for Apache Kafka, Malachyte has helped retailers Google Cloud AI & ML Blog.
The core innovation lies in updating user vectors and predictions in real time—at 100-millisecond intervals—as each behavioral event (click, search, cart action) streams in. Rather than waiting for batch training jobs, every interaction immediately refines the user's profile in Bigtable, while Kafka handles the streaming layer. This multimodal approach encodes browser data, click history, and searches into vectors that move closer to relevant products and further from irrelevant ones. For commerce practitioners, this means achieving far more personalization with less required data or spend than traditional collaborative filtering, while respecting user privacy by reducing dependence on long-term cookie data.
Malachyte's architecture demonstrates that production AI inference requires more than GPUs and storage—it demands real-time continuous learning infrastructure integrating a fast key-value store, streaming layer, and managed messaging system with foundation model architecture. The neural network embedding is continuously trained across all Malachyte retailers, effectively creating a data cooperative where each retailer's user interactions improve the model for everyone.