Ex-Amazon AI Leaders Launch Albatross to Fix What Amazon Can’t
Ex-Amazon AI leaders launch Albatross to replace batch-led personalisation with real-time, transformer-based recommendations across e-commerce – raising €10.5m.
Amazon basically invented modern e-commerce recommendations. The ‘customers who bought this also bought that’ feature became so ubiquitous that it’s hard to imagine online shopping without it. Yet two of Amazon’s own AI leaders just left the company to tear down the entire approach and rebuild it from scratch. Dr Kevin Kahn and Dr Matteo Ruffini raised €10.5 million and launched Albatross from stealth today, betting that even the company that pioneered recommendation systems hasn’t solved the fundamental problem.
Most recommendation systems, including Amazon’s own, still use batch learning. They process data on schedules, looking backward at what you did yesterday or last week. They treat users as static profiles. Albatross claims to be the first platform that learns in real-time, adapting instantly to what you’re doing right now, this second, without manual intervention or retraining. The platform already serves billions of events monthly across marketplaces, retail and travel platforms, processing roughly 100 million products and tens of millions of users.
‘Until now, no platform could adapt instantly to changes in user behaviour,’ said Kahn, who co-founded the company alongside Ruffini and serial entrepreneur Johan Boissard. This is the gap they saw inside Amazon and couldn’t fix from within.
Why Your Recommendations Are Stuck in the Past
Enjoying this article?
Subscribe to our newsletter and never miss a story. No spam, ever.
You’ve searched for hiking boots, so now every website shows you hiking boots for weeks, even after you bought them. Or you browse one item and your feed becomes an endless loop of similar products. This happens because traditional systems rely on historical patterns. Collaborative filtering and content-based systems look at what you did before, what similar users liked or product attributes. They’re fundamentally backward-looking.
Even Amazon Personalise, Amazon’s own recommendation product, faces the batch versus real-time trade-off. According to Amazon’s own engineering documentation, their system combines offline batch collaborative filtering recommendations with real-time clickstream data through complex reranking functions. Batch processing is cost-effective but slow. Real-time updates require careful API management and struggle with cold start problems and data sparsity.
‘Everyone has felt the frustration of seeing the same generic recommendations over and over,’ Kahn said. ‘What’s missing is what truly matters: what the user is doing right now, in this session.’ This challenge of algorithm-driven user experiences extends beyond e-commerce into social media and content platforms.
The Transformer Breakthrough
Albatross uses transformer-based architecture with sequential embedding models trained directly on live events. Transformers are the same architecture behind ChatGPT, but instead of generating text, they’re learning from your browsing sequence in real-time.
Traditional systems treat each interaction as isolated data points. Sequential models understand the order and context. If you looked at running shoes, then yoga mats, then protein powder, the sequence tells a story about your intent right now. The system updates in milliseconds as you browse. Transformer models like BERT4Rec and SASRec use self-attention mechanisms to capture these complex user behaviour patterns, something conventional collaborative filtering simply can’t do.
The research credibility matters. Albatross’s work on cold-start discovery was presented at RecSys 2025, the top conference for recommendation systems, and now powers their production platform. Media streaming service Scribd recently implemented a similar transformer-based approach for real-time recommendations and reported significant improvements in business metrics.
Albatross offers two flagship products: the Real-Time Discovery Feed dynamically curates products and content as you browse, whilst Multimodal Search refines results as your intent evolves. The multimodal capability includes image search that bridges in-store and online experiences. Snap a picture in a shop to personalise your online journey, seamlessly connecting physical and digital retail.
Why They Left
Large companies move slowly, have legacy systems to maintain and face internal politics. Kahn and Ruffini saw the limitation firsthand but couldn’t rebuild the entire recommendation infrastructure from within. Starting fresh gave them freedom to rethink the architecture completely. Amazon pioneered item-to-item collaborative filtering decades ago, but that same success makes it nearly impossible to replace the foundation whilst keeping the business running.
This pattern repeats across big tech. Meta’s Chief AI Scientist Yann LeCun, a Turing Award winner, announced plans to leave the company this month to launch his own startup focused on world models. The rise of proactive AI personal assistants is creating new opportunities that established companies struggle to pursue from within.
The Second Pillar of AI
‘Whilst much of the industry focuses on large language models that generate content, Albatross is building the second pillar of AI: understanding how users perceive and interact with content in real time,’ Kahn said.
Generative AI gets all the attention because it’s flashy. ChatGPT creates essays, Midjourney generates images, Sora produces videos. But understanding and perception AI might be more valuable commercially. When shoppers ask ChatGPT for product recommendations, the AI sees shopping intent and responds with suggestions. This ‘invisible commerce layer’ is becoming critical as AI assistants move into commerce, according to recent analysis from retail analytics firms.
Discovery is becoming the defining challenge as content and commerce explode online. If you can’t help users find what they want instantly, they leave. Albatross reports triple-digit uplifts in engagement from early pilots, though the company hasn’t disclosed which specific metrics (click-through rate, conversion, time on site) or which types of platforms saw these results. Integration takes under seven weeks from signature to deployment, and the platform operates at virtually zero latency with enterprise-grade reliability.
Why Zurich
The founders chose to base Albatross in Baar (Zug) with an office in Zurich. Zurich has quietly become Europe’s AI capital, anchored by ETH Zurich and offices from Google, Meta, Apple and Microsoft. Switzerland allocates 60% of all venture capital to deep tech, the highest share globally according to the Swiss Startup Association’s Deep Tech Report 2025. Deep tech investment reached approximately $1.9 billion in 2024, with projections of $2.3 billion for 2025.
The Swiss tech ecosystem has fostered numerous AI-powered productivity innovations, creating a supportive environment for companies like Albatross to scale their real-time intelligence platforms.
Funding and Traction
The €10.5 million round was led by MMC Ventures, with participation from Redalpine, Daphni and strategic angels. This brings total funding to €13.5 million ($16 million) following a €3 million foundation round in September 2024 led by Redalpine. The team numbers 14 people.
‘Personalisation is entering a new era,’ said Mina Samaan, General Partner at MMC Ventures. ‘Albatross moves beyond static algorithms to build systems that understand context as it happens. Their real-time architecture represents a step change for how businesses engage customers online.’ MMC Ventures has backed other AI-focused companies including Synthesia, the AI video generation platform.
‘We backed Albatross from the very beginning because of the team’s exceptional depth in AI and personalisation,’ said Dr Marc Moesser, Investor at Redalpine. ‘In just a year, they’ve gone beyond what even the largest platforms achieve with real-time infrastructure that adapts instantly to user behaviour.’ Redalpine increased their investment after leading the first round, a strong signal of confidence.
What This Means
If Albatross is right, every marketplace, retailer and content platform will need to rethink their recommendation infrastructure. Static, batch-trained systems will look as outdated as dial-up internet. The question becomes: can you afford to show users yesterday’s recommendations when competitors are adapting in real-time?
The internet has always had a discovery problem: too much content, too little relevance. Search engines helped, then social feeds, then algorithmic recommendations. But all of those still treat you as a static profile. Real-time adaptation could be the next fundamental shift in how we navigate digital spaces. As businesses prepare for AI-powered buyers and automated purchasing decisions, the need for intelligent, adaptive recommendation systems becomes even more critical. Kahn’s vision is to make ‘every interaction intelligent, adaptive and alive.’ Whether that requires leaving Amazon to build it from scratch says something about the limitations of even the most successful tech companies when it comes to reinventing their own foundations.
Agentic AI startup Zalos Raises $3.6M to automate finance and doesn't even need to replace your systems
London and San Francisco-based Zalos has raised $3.6M in seed funding to build autonomous AI agents that log into existing accounting and finance platforms and do the repetitive work humans have been stuck with for years.
