Behavior Intelligence for "Cold Start" Retail
August 5, 2026 • Sidd Motwani
Before Malachyte, my co-founder Ian Anderson and I spent the last decade at Spotify working on the challenge of personalization. At Spotify we built the systems behind its recommendations and made Discover Weekly feel like it knew you, before moving on to tackling search and discovery app-wide and creating dynamic merchandising at scale. There we encountered a unique challenge. The standard methods for doing personalization were manual, could not adapt in real-time, and simply broke at the scale of Spotify's 800M users and more than 1B products.
We soon realized that what was needed was new infrastructure for tracking a user's preferences and making real-time predictions about their intent based on their behavior. At the time, we were confident that, if this “behavior intelligence” infrastructure solved the personalization problem at Spotify, it would ultimately lead to more interesting inventions across a variety of digital use cases.
Today, I'm excited to announce that we've raised $10M in seed funding, co-led by Bessemer Venture Partners and Gradient Ventures, with participation from Harpoon Ventures, to bring this vision to life.
Learning From History: From Steel to Intelligence
We believed we were on the precipice of a new revolution, so we looked to the past for inspiration. In the early 1800s, steel was expensive, brittle, and unreliable. Worse than just being low quality, it was also extremely difficult to produce, requiring days of manual labor. As a result, most steel was produced in small batches for niche applications.
Then came the Bessemer process, one of the greatest breakthroughs of all time that instantly made high-quality steel affordable at scale. It didn't just improve the material. It unlocked the industrial revolution via modern infrastructure: skyscrapers, railroads, and cities.
That's the scale of the shift we're building for today. But, instead of steel, the bottleneck is real-time intelligence, which requires an advanced AI infrastructure that automates big data and understands live behaviors and signals to power discovery, search, and personalization online. While most of the AI conversation right now is about LLMs and agents, the bigger shift is happening underneath that layer in the infrastructure that decides what a system actually knows about a person's intent and how fast it can act on it.
The Attentionless Economy
Today, it's harder than ever for modern brands to truly connect with their customers. Pre-COVID, the tools they've relied on in the digital economy, including search, recommendations, and ad campaigns, have become stale, repetitive, and rigid. For e-commerce brands, the current digital store experience ends up looking the same for everyone. Worse, it's built on assumptions that no longer reflect how people actually shop.
In the post-COVID attention economy, default shopping experiences need to be personal and relevant. If brands can understand and gauge your current intent, curiosity, and need, they will proactively adapt and improve as you browse, not after you come back or purchase something. However, most current technologies can't do this because they were built for the “logged-in web” of yesterday, while most traffic today is anonymous.
Furthermore, old-school “personalization” guesses from past purchases and relies on pre-defined rules. For example, if a shopper previously bought a onesie for an infant, the personalization rule-set assumes they must be a new parent, ignoring the fact that none of their other behaviors signal this to be true; when in reality, that purchase may have been a baby shower gift for a co-worker.
In the attention economy, relevance has to come from live intent — what a shopper is signaling right now and how those intentions evolve in real time. Like the Bessemer process did for steel, modern personalization requires a new infrastructure in order to deliver outcomes in the attention economy. The new infrastructure must:
- Rank real intent instead of guessing from stale purchase history
- Offer a transparent merchandising layer instead of an opaque ad auction
- Be modular and easy to test and adjust instead of locked in
We call this behavior intelligence because it's what happens when a system stops treating past purchases as the whole story. Here's how it stacks up against the recommendation systems most brands run today:
| Dimension | Legacy Recommendations | Behavior Intelligence |
|---|---|---|
| Training cadence | Batch — nightly or weekly retrain | Continuous learning — minute-level parameter sync |
| Cold start | Rules or heuristics; days to weeks to “warm up” | Solved via representation fusion in seconds |
| Transfer across tasks | One model per surface | One representation, every surface |
| Serving latency / scale | Varies; rarely engineered for real-time | Sub-200ms; tens of millions of queries/sec |
Malachyte's solutions for search, recommendations, and product pages are based on behavior intelligence, so that every online experience can proactively adapt to shoppers in real time while simultaneously future-proofing enterprises for the attention economy. Training the model on user behavior allows us to build a generalized understanding of intent, then layer in continuous learning infrastructure so that understanding adapts in real time and keeps evolving without ever needing to be retrained. This allows brands and retailers to keep driving results even as what people want, and when they want it, keeps changing.
Our Vision for Modern Commerce
People aren't what they purchased once on a whim. People are the sum of all of their nuances. Instead of batch-training on stale data and hoping the patterns still hold, our behavior intelligence learns continuously, updating its read on intent from the signals of real users in real time. Everything they're doing right now (clicking, hovering, comparing, abandoning) is valuable.
There's no cold start, no waiting for enough history to make a confident guess, no retraining pipeline running on a weekly cron job. There are no third party cookies invading your privacy or permissions pop-ups crowding the screen and disrupting your experience enough to make you close the browser out of frustration. It's a system of self-learning that translates to good predictions that are built from this moment, not the last quarter.
As a user, just imagine:
- Product discovery that evolves based on your curiosity
- Search that understands what you mean, not just what you typed
- Recommendations that feel intuitive and contextually relevant, not just the same old ‘people similar to you also like’ we've become accustomed to
This allows brands to show up differently, yet meaningfully, for every visitor enabling them to identify and act on emotional signals and intent without needing prior data or manual rules. It's not a chatbot, but the real-time operating system beneath discovery, search, and personalization. This is personalization in the AI era and it's based on the same vector AI math that Ian and I used at Spotify to drive recommendations.
Built by a Team That's Done This at The Frontier
We started Malachyte because too many AI systems reduce people to clicks and cohorts. Our promise is to put humans at the center: recognizing momentary intent without hoarding identity, and giving brands transparent, ethical control over how decisions get made.
Alongside fellow co-founder and Chief Operating Officer, Shivaditya Sinha, Ian and I have built a founding team of AI experts who have firsthand experience with scale, consumer behavior, and some of the largest commerce operations in the world. Every week, we're signing new and visionary brands and retailers who understand the full potential of reimagining personalization and product discovery to drastically improve the shopping experience from the ground up.
If you lead digital, marketing, or customer experience at a brand or retailer, I invite you to join us on this journey.
