RecipeBook lets AI teams rank and purchase video datasets in minutes
The YC-backed startup says users can re-rank 25 million videos and purchase selected clips for $3 per hour.
By Ryan Merket · Published
Why it matters
RecipeBook packages video-data procurement into a search, feedback and purchase loop. If Shofo's speed and pricing hold at scale, small AI teams could build narrow datasets without a long managed-data contract.
Braiden Dishman (@BraidenDishman), a co-founder of Shofo, launched RecipeBook on July 27th, giving AI developers a self-service interface for searching, ranking and purchasing video training data. Dishman said in a thread on X that the initial corpus contains more than 25 million videos and sells for $3 per selected hour.
The release turns part of Shofo's custom data operation into a storefront. A user enters a search, rates relevant clips, trains what the interface calls a "probe," re-ranks the results and adds a chosen number of top-ranked hours to an order. The live application includes filters, ordering controls and buttons to clear ratings or retrain the probe.
Dishman said data vendors had quoted prices ranging from $15 to $480 per hour, with delivery measured in weeks or months. Shofo claims RecipeBook can deliver selected data within minutes. Those price and delivery comparisons come from Shofo, and the company has not published the vendors or specifications behind them.
From social search to video data
Shofo is a four-person San Francisco startup founded in 2025 and backed by Y Combinator through its Winter 2026 batch. YC launch materials identify Bryan Hong as CEO, Dishman as COO, Alexzendor Misra (@AlexzendorMisra) as CTO and Andre Braga (@AndreBragaML) as head of AI.
The founders met while building Correkt, a multimodal AI search engine. The Daily Nexus reported that Correkt launched in April 2024 as a citation-focused search product and attracted about 5,000 users shortly after release. Shofo's YC materials say Correkt eventually passed 40,000 users before the founders changed direction.
Hong described the pivot in a December 2025 post: the founders shut down Correkt, turned down internships at Stripe and Meta, and moved into a $1,000 room in San Francisco to build live indexes of social platforms. Dishman's YC profile lists an economics background at the University of California, Santa Barbara and prior work at Amazon Web Services. Misra previously led Correkt, while Braga studied statistics and data science at UCSB and worked on reinforcement learning as a 2025 MIT Summer Research Program intern.
Shofo initially presented itself as an API for live social data. Its current pitch focuses on assembling video datasets for AI labs from an index that Shofo says contains billions of videos. The Shofo website advertises custom collections drawn from millions of hours of video, with footage cleaned, segmented and labeled for model training.
RecipeBook narrows that pitch into a buyer-controlled workflow. Shofo's main site still directs customers to request a custom dataset, while RecipeBook lets researchers conduct the initial search and relevance-ranking work themselves. That can reduce the costly back-and-forth involved in defining a dataset, particularly when a model developer needs a narrow visual concept or interaction rather than a general-purpose video collection.
The contract behind the clips
Shofo says its larger index combines public videos collected from the open web with private videos aggregated from thousands of sources. Its enterprise terms grant customers a perpetual, worldwide license to use purchased datasets for training, fine-tuning, evaluation, benchmarking and model development. The license is non-exclusive and generally non-transferable.
The same terms provide the service "as is" and disclaim warranties including title and non-infringement. RecipeBook also maintains a copyright removal process for rights holders. Buyers evaluating the $3-per-hour offer will still have to weigh provenance, licensing terms and dataset quality alongside procurement speed.
RecipeBook is a test of whether Shofo can make highly specific video data behave like searchable cloud inventory. The founders already built one multimodal search product and then repurposed that indexing experience for AI infrastructure. With RecipeBook, they are betting that researchers want direct control over the final mile: deciding which examples belong in a training set before paying for the underlying video.