Three Ex-Spotify Engineers Build Real-Time Intent Engine for Online Retail

From Music Streams to Shopping Carts

Three engineers who spent years refining the behavioral intelligence layer behind Spotify's recommendation system have turned their attention to a problem they believe plagues digital commerce: most online stores still treat shoppers as if they're reading from yesterday's script.

Sidd Motwani, Ian Anderson, and Shivaditya Sinha built Vector AI at Spotify, the infrastructure that now drives roughly 90% of recommendations to the platform's 800 million users. The system distinguishes itself by attempting to predict intent and next actions rather than simply extrapolating from past behavior. Their new venture, Malachyte, applies that same philosophy to e-commerce, where personalization engines typically lean heavily on purchase history, demographic buckets, or logged-in profiles.

The startup announced it has secured $10 million in seed funding co-led by Bessemer Venture Partners and Gradient, with Harpoon Ventures participating. The capital will fund distribution expansion and bring in product and commercial leadership.

The Problem with Purchase-History Personalization

At DailyTechWire, we've tracked how recommendation systems evolved across Asia's e-commerce giants, from Alibaba's collaborative filtering to Coupang's warehouse-proximity algorithms. Yet a common constraint persists: most platforms optimize around what a customer bought last month, not what they need this afternoon.

Malachyte's thesis is that first-time visitors land on generic storefronts identical to everyone else's, while returning customers see suggestions anchored to their order history rather than their current task. A shopper who bought running shoes six months ago may now be furnishing an apartment, but the homepage still surfaces athletic gear.

According to Motwani, now CEO of Malachyte, retailers already capture the richest signal stream available but rarely act on it in the moment. Every hover, scroll, search refinement, and cart addition carries information about preference and immediate intent. Most systems either ignore these micro-actions during the session or batch them into overnight segments for the next day's targeting.

Two-Headed Vector Architecture

Malachyte's platform uses what the team calls a two-headed Vector AI design. One vector tracks general taste, the stable preferences that persist across sessions. The other captures session-specific intent, the immediate goal driving behavior right now.

The system begins forming a profile before the first click, using context available at page load: traffic source, device type, time of day, geographic signal. A visitor arriving from an email link on a phone at 11 p.m. occupies a different mental state than the same person on a laptop mid-morning, yet conventional systems treat both sessions identically.

Within a single session, Malachyte claims it can build a actionable read on both dimensions. A search for "heavy-duty boot" followed by two clicks on steel-toed models is sufficient to elevate work pants and gloves on the page while demoting dress shoes, no account or purchase history required. Each subsequent action sharpens both vectors, so relevance compounds the longer someone stays and carries forward to the next visit.

This approach mirrors the real-time tuning that powers Spotify's Discover Weekly and Daily Mix playlists, where the algorithm balances long-term taste with short-term mood. In music, that might mean surfacing upbeat tracks during a morning commute and ambient instrumentals late at night. In commerce, it translates to prioritizing task-relevant products over affinity-based suggestions.

Testing Ground and Shopify Integration

Malachyte has been in development since 2024, working with more than 20 enterprise clients across travel, grocery, and retail before narrowing its focus to e-commerce. The platform went live in production for the first time in fall 2025 with Fun.com, an online retailer specializing in pop culture merchandise and costumes.

Since June 2026, the technology has been generally available to Shopify merchants through a native integration. Larger retailers outside the Shopify ecosystem can access the system via API, allowing integration with custom front-ends and content management platforms.

The Shopify route is strategic. The platform hosts more than two million active stores, many of them small and mid-market merchants without dedicated data science teams. Malachyte's plug-and-play model lowers the barrier to real-time personalization, a capability historically reserved for companies with the engineering resources of an Amazon or Zalando.

For enterprise retailers, the API path offers more flexibility but requires heavier implementation lift. The calculus hinges on whether the incremental conversion gain justifies the integration cost, a question that will likely hinge on basket size and repeat-purchase frequency.

Contextual Signals and the Next Layer

Motwani argues that contextual signals remain significantly underutilized in digital commerce. Device type, referral source, time of day, and even scroll velocity all carry predictive weight, yet few platforms incorporate them into decisioning in real time.

The insight aligns with broader trends across recommendation infrastructure. In Seoul, Naver's shopping vertical has experimented with incorporating weather data into product ranking, surfacing umbrellas and rain jackets when precipitation is forecast. In Singapore, Shopee tests dynamic homepage layouts based on commute patterns, shifting from browse-optimized grids during lunch breaks to conversion-focused carousels in the evening.

Malachyte's ambition extends beyond on-site personalization. Motwani sees the larger opportunity in unifying merchandising and marketing around a shared understanding of customer behavior. Today, those functions often operate in silos: merchandising teams curate collections and set promotional priorities, while marketing teams run segmented campaigns based on separate customer data platforms. A common vector layer could allow both to react to the same real-time signal, aligning homepage placements, email content, and paid acquisition creative around a single view of intent.

That convergence is still ahead. For now, Malachyte's product focuses on on-site experience, where latency and integration complexity are more manageable.

Funding and the Road Ahead

The $10 million seed round positions Malachyte to scale its go-to-market motion and expand its product and commercial teams. Bessemer Venture Partners, which backed Shopify, Twilio, and Toast, brings a track record in infrastructure and vertical SaaS. Gradient, Google's AI-focused venture arm, adds domain expertise in machine learning systems and model deployment.

The funding environment for recommendation and personalization startups has tightened since the 2021 peak, when companies like Bloomreach and Dynamic Yield commanded nine-figure valuations. Investors now scrutinize unit economics more closely, favoring platforms that can demonstrate measurable lift in conversion or average order value within weeks of deployment rather than months.

Malachyte will need to prove that its real-time approach delivers enough incremental revenue to justify the switching cost from incumbents like Nosto, Algolia, or in-house systems. Early traction with Shopify merchants offers a testing ground, but enterprise adoption will require case studies showing sustained performance across categories and geographies.

The broader question is whether intent prediction at the session level can overcome the cold-start problem that has long constrained recommendation engines. Spotify benefits from explicit feedback signals: skips, saves, and playlist adds. E-commerce lacks that same richness. A click on a product could indicate interest, comparison shopping, or accidental navigation. Cart additions don't always convert. Returns muddy the signal further.

Malachyte's bet is that volume and velocity of micro-actions compensate for signal ambiguity. If a shopper's sequence of behaviors within a single session carries enough information to infer intent, the system can bypass the need for deep historical data. Whether that holds true across product categories, price points, and purchase cycles will determine how far the technology scales beyond its initial beachhead.