Voxel51 Inc., a platform that helps visual artificial intelligence model developers curate and refine their data to increase the accuracy of their AI models, today announced that the company has raised $30 million in new funding to expand its products.

The Series B investment was led by Bessemer Venture Partners, with participation from new investor Tru Arrow Partners and existing investors Drive Capital, Top Harvest Capital, Shasta Ventures and ID Ventures. The company said that it will use the new funding to expand its community, invest in AI research and boost its market reach to help AI businesses in the demanding market of visual-oriented AI.

Co-founder and Chief Executive Brian Moore told SiliconANGLE in an interview that there’s a paradox with visual data that most bits going through routers on the internet right now are visual but insights from that data go largely untapped. Although most AI companies gravitate toward the best and the brightest foundation models, when it comes to visual data, that’s not the best way to increase accuracy and lower error rates.

“In practice, the difference between success or failure from these projects came from data quality, and not necessarily model architectures,” said Moore. “You could always take the latest and greatest off the shelf, but the reason that 80% of data projects are not reaching production is usually a data quality issue.”

Voxel51’s flagship products include the open-source tool FiftyOne and the enterprise-scale service FiftyOne Teams, which provide the building blocks for optimizing a visual dataset pipeline and analyzing its labeling, scenarios of interest, identifying failure modes and determining annotation mistakes. With the toolset, data engineers have a fully-fledged toolset for sifting through visual data to refine and evaluate model behavior that can assist them in refining AI models to make sure that there aren’t any hidden issues.

Both products also provide access to what the company calls FiftyOne Brain. It’s a library of machine-learning capabilities that can provide recommendations and insights into a visual dataset for users to guide them on how they can get better performance for their models. The Brain can be used to diagnose failure cases in visual AI models when an AI can’t understand the difference between visual embeddings when there are annotation mistakes or other issues within the dataset.

Using VoxelGPT, an AI chatbot added in June, users can ask natural language questions about their visual datasets. The chatbot can provide images and code samples, give answers regarding documentation, tutorials and application programming interface queries in English. This makes it easier for developers to dive into their datasets or get moving with their workflows.

Both FiftyOne and FiftyOne Teams also have native vector search integrations with popular databases, including Pinecone, MongoDB, Redis, Qdrant and more, making it possible for fast and easy searches through billions of images.

Moore explained that these tools are foundational for the future of AI model development, especially the advancement toward AGI, or artificial general intelligence, the holy grail of AI that is aimed at producing models that can perform tasks with human-like cognition. Large language models can appear to seemingly reason and converse like humans in text and voice, but AI will also have to interact visually and understand the physical world for that to happen.

This has played out recently with big visual AI announcements from OpenAI with its GPT-4o model, which has visual capabilities, and Google LLC’s Project Astra, capable of chatting and reasoning about video. Both of these product announcements, within days of each other, Moore said, show that the industry is hot on the trail.

“It’s top of mind how folks can tap into visual data,” said Moore. “And obviously, there’s been high-profile projects underway.”

Image: Voxel51