He built the video models that help pick your Instagram Reels. Then he left Facebook to sell that machinery to companies that could never build it themselves.
Tullie Murrell. The face behind a library that decides, in milliseconds, what millions of people watch next - and a startup betting that anyone should be able to do the same.
Shaped is a retrieval engine for search, feeds, and agents. It sits between a company's data and its users, decides what to show and in what order, and returns an answer in milliseconds. This is the kind of infrastructure Facebook, Netflix, and TikTok spend years and hundreds of engineers building. Tullie Murrell's argument is that a five-person team should be able to rent it instead. He has some credibility on the point: he helped build one of the famous versions.
At Facebook, Murrell led PyTorchVideo, the domain library that supplied the models and datasets behind several video-understanding products, including newsfeed ranking and Instagram Reels. The obvious startup, if you were him, was a video-personalization company. He talked himself out of it.
"It's not the video personalization that really matters here," he told TechCrunch. "Why don't we sell a horizontal personalization platform?" That single reframing is most of Shaped's strategy. The company treats users, content, and events as embeddings, learns from them continuously, and exposes the result through an API and a query language called ShapedQL.
The convenient part, Murrell notes, is that once you understand users and content semantically, search comes almost for free. "We have all this semantic understanding about the users and content, so it's actually quite easy for us to add on the search packaging." Two products that used to require two teams collapse into one system.
A real-time retrieval engine for search, feeds, and agents. Connect data from Snowflake, BigQuery, Databricks, Postgres and more; pick a model like Llama, CLIP, or BERT; retrieve ranked results in milliseconds. Based in New York, roughly 27 people.
Outdoorsy improved search relevance with real-time personalization. Trela, a Brazilian grocery-delivery service, reported a 16% increase in average order value. QVC is a named customer.
ShapedQL is a domain-specific dialect for building ranking pipelines. It handles four stages - candidate retrieval from multiple sources, filtering with hard constraints, ML-based scoring, and reordering for diversity - in a form that reads like SQL. The pitch to engineers is not that Shaped does something they cannot. It is that Shaped does the boring 2,000 lines so they do not have to.
"We're not just building another personalization engine. We're building the foundation for a future where every user interaction is informed, intelligent, and impactful."
Tullie Murrell / on Shaped's Series A
"How do we help small businesses build the same kind of hyper-personalized experiences that you'd see in a big tech company?"
"It's not the video personalization that really matters here. Why don't we sell a horizontal personalization platform?"
"We have all this semantic understanding about the users and content, so it's actually quite easy for us to add on the search packaging."
Murrell studied computer science at the University of Adelaide and interned at Palantir, Google, and Uber before graduating. He met his future co-founder, Daniel Camilleri, at a hackathon in Australia about a decade before they started a company together. The two problems - build the technology, and find the person to build it with - were both solved early.
His open-source work on PyTorch Lightning made him an Arctic Code Vault Contributor, which is to say a copy of his code is archived on film in a decommissioned coal mine in Svalbard, meant to survive a very long time. The GitHub badge is a small, strange trophy for a career spent making machine learning legible to other people.
BSc in Computer Science. Interned at Palantir as a forward-deployed engineer, and at Google and Uber as a software engineer.
Roles across FAIR, Building 8, and ML engineering.
Builds the leading video-understanding library and contributes to PyTorch Lightning. Models feed newsfeed ranking and Instagram Reels.
Leaves big-tech research to democratize AI-powered personalization with Daniel Camilleri.
Shaped goes through the winter batch.
Led by Madrona Ventures; self-serve platform launches. Karan Mehandru joins the board.
The leading open-source library for video-understanding research and production. Supplied models and datasets used across Facebook's video products.
Standardized how deep-learning models are trained. Now used by millions of developers who mostly do not know his name.
Video understanding, 3D vision, and medical imaging. A tour of exactly the domains where getting data into a model is the whole battle.
Madrona Ventures met Murrell more than a year before leading the round and came away impressed by his understanding of recommendation systems. Their write-up used the phrase venture investors reserve for people who have spent years obsessing over the exact problem they now sell: perfect founder-market fit.
The $8M Series A closed in July 2024 with participation from Y Combinator and operators from Clickhouse, Docusign, Okta, Rippling, and StitchFix - a cap table heavy on people who have themselves built or bought this kind of infrastructure. Total funding to date is around $10M.
The video models he led at Facebook help decide which Instagram Reels you see.
ShapedQL can replace roughly 2,000 lines of ranking maintenance code with about 30.
He is an Arctic Code Vault Contributor - his open-source code is archived on film in a Svalbard mine.
He interned at Palantir, Google, and Uber before finishing his undergraduate degree.
He met co-founder Daniel Camilleri at an Australian hackathon about a decade before Shaped existed.
One of his pinned GitHub repositories is just his own Vim configuration files.
Tullie Murrell is the CEO and co-founder of Shaped, a New York startup building a real-time retrieval engine for search, feeds, and agents. Before Shaped, Murrell spent about five years at Facebook AI Research, where he created PyTorchVideo, the leading open-source library for video understanding, and was an early core contributor to PyTorch Lightning, a training framework now used by millions of developers. The video models he led powered Facebook newsfeed ranking and Instagram Reels. In 2021 he left big-tech research to package that same personalization machinery into a self-serve product for companies without armies of ML engineers. Shaped went through Y Combinator's W22 batch and raised an $8M Series A led by Madrona Ventures in July 2024.
Last updated: