Malachyte uses Google Cloud to tackle eCommerce cold start

Malachyte has built an eCommerce recommendation system on Google Cloud to help retailers address the cold start problem for first-time visitors.

The platform uses attention-based neural networks to predict what a shopper is likely to want next from the sequence of actions taken during a browsing session. Malachyte applies the same model to on-site search, product pages, category pages and add-to-cart recommendation slots.

At the centre of the approach is a user vector created when a visitor arrives on a retailer's site. At that point, little is known about most shoppers, making personalisation difficult.

Each click, search and other interaction updates the user vector while generating a new prediction. This lets the system refine its understanding of an individual shopper during a session while also improving the broader model with anonymised data.

Recommendation and search agents refresh results for the next page in about 100 milliseconds based on actions taken on previous pages. That speed is intended to let retailers change search results and recommendation carousels without interrupting the user experience.

Sales impact

Malachyte said some retail customers have doubled, and in some cases tripled, sales after adopting the system. It did not say how many retailers were included in those results.

The model is designed to rely less on long-term profiles and third-party cookie data than older collaborative filtering systems. Instead, it focuses on behaviour within a live session, which Malachyte argues can deliver more relevant recommendations with less historical data.

Products in a retailer's catalogue are embedded in the same vector space as the user profile. As a shopper interacts with a site, the user vector shifts closer to products judged relevant and further from those considered less relevant.

The neural network behind those embeddings is trained continuously across retailers using the platform. That means data from one retailer's users can help improve the shared model, though the data used for model improvement is anonymised, according to the company.

Cloud stack

For the infrastructure, behavioural events stream into Managed Service for Apache Kafka on Google Cloud. Those events are then used to update user profiles stored in Bigtable, rather than waiting for a later training batch.

Kafka and Bigtable each operate at roughly 10 milliseconds per step within the recommendation loop, according to the company. That timing enables the wider system to complete inference and profile updates quickly enough to support real-time changes on a retail site.

A second layer handles product catalogue changes, inventory signals and retailer-specific data through Cloud Pub/Sub. Malachyte said the service's global REST APIs allow retailers to connect product feeds without heavy integration work.

The company runs its agents on Google Kubernetes Engine, while model inference runs on Google Compute Engine. The broader setup combines streaming, messaging and key-value storage with the recommendation model, rather than treating training and serving as a stand-alone GPU task.

Industry challenge

The cold start problem has long been a central challenge in digital commerce because retailers often know little about a visitor when they first arrive on a website or app. Older recommendation systems commonly depended on large stores of interaction history, demographic information or cookie-based tracking to identify similar users and products.

Malachyte's approach instead treats a shopping session as a sequence, much as language models treat words in a sentence. By focusing on order and context across a session, it aims to infer intent from limited signals.

Sidd Sinha, founder of Malachyte, said the company was created after work on recommendation systems for larger businesses including Spotify and Priceline. He said the goal was to apply techniques associated with large language models to established personalisation problems in retail.

Google Cloud used the example to argue that production AI systems require more than computing power and storage. In Malachyte's case, the architecture depends on continuous learning infrastructure that can ingest events, update user context and return new recommendations during an active shopping session.

Malachyte said the result is a system that can personalise search and recommendations for users it knows little about while reducing reliance on longer-term tracking data.