Each team built its own stack; every request reprocessed thousands of images. A shared platform with precomputed embeddings cut latency 44x.
One of the world's largest visual discovery platforms, with 640 million monthly active users, 80 billion+ monthly searches, and a Taste Graph built from 16 billion+ boards, where more than 96% of text searches are unbranded.
Expanding into vision-language models created a fundamentally different infrastructure challenge: a single production request may carry thousands of...
“Building the next generation of AI-powered discovery means investing in infrastructure that can keep up with the scale and complexity of Pinterest. Our collaboration with NVIDIA helps us deliver faster, smarter and more personalized experiences for the hundreds of millions of people who use Pinterest.”
Visual discovery engine for recipes, home inspiration, and style ideas.
NVIDIA is a technology company that specializes in semiconductors, graphics processing units, and artificial intelligence for applications in data centers, gaming, and more.
Pinterest's Visual search is part of this use case:
Related implementations across industries and use cases
Moderation couldn't keep pace with 600M users. AI agents now filter toxicity while models recognize 2.5B objects to refine search.
Elasticsearch treated creator vectors like log data, aging 400M+ to frozen storage, pushing search to 25s. A vector-native store fixed it.
Messy ingredient data broke text search, capping matches at 20%. AI semantic search now maps free-form inputs to unlock recipe imports.
Moderation couldn't keep pace with 600M users. AI agents now filter toxicity while models recognize 2.5B objects to refine search.
Elasticsearch treated creator vectors like log data, aging 400M+ to frozen storage, pushing search to 25s. A vector-native store fixed it.
Sequential AI testing bottlenecked development. Engineers built a concurrent, code-first pipeline to evaluate agent responses in seconds.
Reaching a new market meant standing up another dubbing team. One API now dubs into 30+ languages; creators edit tracks before it goes live.
Tournaments running simultaneously meant an hour of manual checks each. AI agents now run them in minutes, freeing the team to be proactive.
Large AI training jobs meant fighting for preemptible slots or leaving campus. Marlowe gave any lab guaranteed multi-node access on demand.
Each team built its own stack; every request reprocessed thousands of images. A shared platform with precomputed embeddings cut latency 44x.
One of the world's largest visual discovery platforms, with 640 million monthly active users, 80 billion+ monthly searches, and a Taste Graph built from 16 billion+ boards, where more than 96% of text searches are unbranded.
Expanding into vision-language models created a fundamentally different infrastructure challenge: a single production request may carry thousands of...
“Building the next generation of AI-powered discovery means investing in infrastructure that can keep up with the scale and complexity of Pinterest. Our collaboration with NVIDIA helps us deliver faster, smarter and more personalized experiences for the hundreds of millions of people who use Pinterest.”
Visual discovery engine for recipes, home inspiration, and style ideas.
NVIDIA is a technology company that specializes in semiconductors, graphics processing units, and artificial intelligence for applications in data centers, gaming, and more.
Pinterest's Visual search is part of this use case:
Related implementations across industries and use cases
Moderation couldn't keep pace with 600M users. AI agents now filter toxicity while models recognize 2.5B objects to refine search.
Elasticsearch treated creator vectors like log data, aging 400M+ to frozen storage, pushing search to 25s. A vector-native store fixed it.
Messy ingredient data broke text search, capping matches at 20%. AI semantic search now maps free-form inputs to unlock recipe imports.
Moderation couldn't keep pace with 600M users. AI agents now filter toxicity while models recognize 2.5B objects to refine search.
Elasticsearch treated creator vectors like log data, aging 400M+ to frozen storage, pushing search to 25s. A vector-native store fixed it.
Sequential AI testing bottlenecked development. Engineers built a concurrent, code-first pipeline to evaluate agent responses in seconds.
Reaching a new market meant standing up another dubbing team. One API now dubs into 30+ languages; creators edit tracks before it goes live.
Tournaments running simultaneously meant an hour of manual checks each. AI agents now run them in minutes, freeing the team to be proactive.
Large AI training jobs meant fighting for preemptible slots or leaving campus. Marlowe gave any lab guaranteed multi-node access on demand.