Livestream shopping app Whatnot acquired Shaped, a machine learning company built around real-time recommendation and search systems, the company announced Wednesday. TechCrunch AI reports the deal brings Shaped founder and CEO Tullie Murrell plus nearly a dozen engineers and AI researchers into Whatnot, where Murrell will run a newly formed Applied AI Research group. Terms weren’t disclosed.
This is an acqui-hire aimed squarely at a problem most e-commerce companies never have to solve.
Why live commerce breaks normal recommendation engines
Standard e-commerce recommendations work because catalogs sit still. Amazon knows what’s in stock. The product page you saw yesterday is the same one today. A model trained overnight is still useful in the morning.
Live auctions destroy that assumption. Inventory changes by the second. Shows start and end continuously. A buyer who wanted vintage cards ten minutes ago is now watching a sneaker drop.
“By combining Shaped’s technology with Whatnot’s existing systems, we can make recommendations faster, more responsive, and more personalized,” Emmanuel Fuentes, VP of Data and AI at Whatnot, told TechCrunch. “That speed matters because live commerce is a uniquely hard recommendation problem. Inventory changes by the second, shows start and end continuously, and buyer intent shifts throughout a show.”
What stands out here is the latency number. Fuentes said Whatnot spent six years cutting recommendation lag from roughly a day down to minutes. Shaped is supposed to close the remaining gap to something near real time. That’s the whole bet.
The scale behind the problem
The numbers explain why Whatnot bought instead of built:
- 500,000+ hours of live video processed weekly
- Millions of real-time interactions per week feeding the models
- 1 billion+ orders placed by sellers since the 2019 launch
- 20 million buyers added in the past year
- $225 million Series F last year, valuing the company above $11 billion
- 80+ new categories launched since last year, including art, golf, and vinyl
Every new category makes discovery harder. When you sell trading cards, matching buyers to sellers is straightforward. When you sell trading cards and vinyl and golf clubs across thousands of simultaneous live shows, it becomes a genuine machine learning problem.
What Shaped actually built
Shaped combined customer data with large language models and machine learning to power personalized search and discovery. Its clients included Outdoorsy and QVC, per TechCrunch AI. QVC is the tell: a company that’s been doing live selling on television since 1986 was paying Shaped to modernize how it surfaces products.
Murrell came from Meta before founding Shaped. That background matters for context. Meta’s recommendation infrastructure is the reference implementation for real-time ranking at scale, and Whatnot just hired someone who worked inside it.
The bigger pattern
Whatnot isn’t moving alone. eBay and Poshmark are both racing to push AI through their platforms, and resale is turning into a discovery arms race. The platform that matches the right buyer to the right live show fastest captures the sale. Everyone else shows you something you already scrolled past.
Three things worth watching:
- Acqui-hires are the default AI staffing move now. A dozen specialized ML engineers are nearly impossible to hire individually. Buying the company they already work at is faster.
- Real-time is the new competitive line. Batch recommendation systems are becoming table stakes. Latency is where the differentiation lives.
- Shaped’s existing customers are in limbo. Outdoorsy and QVC now depend on technology owned by a company with different priorities. Expect migration announcements.
If you’re building recommendation systems, the lesson is direct: the ceiling isn’t model quality anymore, it’s how fast you can serve a prediction against data that’s already changed. Whatnot just paid to move that ceiling.
More details on the acquisition are available at the original source.