Startup Lovelace targets contextual AI engine at mission-critical use cases

Lovelace AI Inc. is emerging from stealth mode today with an approach to enterprise artificial intelligence that it says is necessary for high-stakes decision-making, particularly in environments where errors can have significant consequences.

At the core of the company’s approach is a “context engine” builder called Elemental that’s designed to sit between AI agents and underlying data systems. Elemental creates secure, enterprise-specific context engines that create structured knowledge graphs from fragmented data that AI agents can navigate and query to return research-quality analysis with citations.

The back end, called YottaGraph, can handle trillions of interconnected facts to augment internal data with intelligence from external sources.

Lovelace AI, which is named after Ada Lovelace, the 19th-century mathematician widely recognized as the first computer programmer, was founded by Andrew Moore, former head of AI in Google LLC’s cloud division, dean of Carnegie Mellon University’s School of Computer Science and the first AI adviser for the United States Central Command within the Department of Defense.

Moore said he conceived the idea for Elemental while at Google. “I mostly worked on placing big artificial intelligence systems inside big enterprises like banks, hospital systems and manufacturing plants,” he said. “I did not see a path to being able to use the summarized chat that you use with large language applying to big, serious enterprise AI problems.”

Elemental creates secure, enterprise-specific context engines that transform fragmented data into structured knowledge graphs that AI agents can quickly navigate and query. “When an AI agent is asked to do an investigation, instead of having it talk directly to this huge repository of data, we insert something between that figures out exactly what root sources of data are appropriate for answering the question,” Moore said.

Over the two years it has been in stealth, Lovelace has primarily targeted industries where lives are at stake, such as public sector agencies, national security, disaster response and healthcare. However, Moore said, the technology is also appropriate for high-risk, high-reward scenarios such as financial services.

Moore said Yottagraph can answer investigative questions with about one 1,000th of the token use of other methods. “With 1,000th of the tokens, you can afford to ask 1,000 questions where you might have only previously been able to afford one,” he said.

Lovelace is currently ingesting about a billion facts a week from about 20 different public sources, he said. A key differentiator is how the system handles relationships between data points. “We shunt millions and sometimes billions of little messages between nodes in this knowledge graph,” Moore said. “That does the computation in a way that’s much more suited for CPUs, GPUs and the memory layout of database systems.”

Sources include “all the news in the world, all the social media in the world,” as well as shipping data, logistics information and satellite imagery.

Entity resolution is critical to ensuring accuracy by “making sure that irrespective of the source, we know the shared identity of the central entities at play,” he said. “You can’t afford to have errors in your entity graph.”

To maintain trust, the system tracks provenance for every inference so users can identify every factor in reasoning. “That means that whenever we do a large chain of reasoning, we can give citations to every part of the process,” he said.

The platform is deployed within customer environments rather than as a centralized service. Providing that level of control is essential in sensitive environments, Moore said. “Gone are the days when a vendor can say, ‘Send me your data, I’ll process it for you and send you back the results,’” he said.

The company will focus primarily on large enterprises, particularly in scenarios “where executives are under pressure to demonstrate tangible AI-driven productivity gains,” Moore said. Those include areas where “actual implementations have failed far more often than they’ve succeeded.”

Image: SiliconANGLE/Microsoft Designer