Activeloop Deep Lake

Activeloop operates as Snark AI, Inc., a company with a Mountain View, California address in its terms, and says it began building Activeloop in 2018. A March 2024 post reports an 11 million USD Series A from investors including Y Combinator and Samsung Next. It sells infrastructure that stores and serves data for AI agents and model training.

Its Deeplake database combines vector and tensor storage, GPU streaming for fine-tuning and a serverless Postgres interface queried through SQL. Hivemind captures agent traces as shared team skills, and Refinery supports continuous improvement loops. Named customers include Bayer, Matterport and Flagship Pioneering. In December 2025 the company open-sourced Deep Lake PG and the C++ code behind Deep Lake.

• Vector and tensor storage: embeddings and multimodal data in one store.
• GPU streaming: feeds data to GPUs for model fine-tuning.
• Serverless Postgres interface: SQL queries over multimodal data via Deep Lake PG.
• Hivemind: captures agent traces and shares them across teams as skills.
• Refinery: feedback-driven loop for continuous improvement.
• Open source: C++ core and Deep Lake PG published on GitHub in December 2025.

• A Vietnamese AI company managing image and video datasets for model training.
• An enterprise building a RAG assistant over documents with embeddings and metadata in one store.
• A medical-imaging or manufacturing inspection team streaming large datasets to GPUs.

• Vector and tensor data live in one store, which avoids stitching separate databases together.
• Open-source code lowers lock-in and allows self-hosting.
• SQL access through serverless Postgres fits existing data teams.
• Case studies report up to 80% less data preparation time at Matterport and 18% more accurate RAG at Flagship (vendor claims).

AI agents and multimodal models need storage for embeddings, images and traces alongside transactional state, and teams look for ways to avoid running several specialised databases. Vietnamese companies building computer-vision, document-AI or RAG products can use a unified vector and tensor store for training and retrieval data.

Operates as Snark AI, Inc. dba Activeloop, Mountain View, California (per terms). Began building Activeloop in 2018. Raised an 11 million USD Series A (announced March 2024) from Streamlined Ventures, Y Combinator, Samsung Next, Alumni Ventures and Dispersion Capital.