Startup tackles knowledge graphs to improve AI accuracy

The lack of reliable context has become a critical barrier to the adoption of AI agents, but startup Lovelace says it has solved the problem.

The answer, said co-founder Andrew Moore, the former head of Google Cloud AI, is his company’s new platform, Elemental, an AI-powered system for building knowledge graphs that he said is cheaper, faster, and more accurate than its competitors. It can help ground large language models (LLMs) in accurate context, he explained, while also providing full auditability so that enterprises know the exact information on which decisions were made.

“You cannot do safety-critical reasoning for agents purely based on trying to do the same sort of thing you normally do for chatbots,” Moore told CIO. “You need something else to help make sure that the reasoning that’s going on is properly coordinated.”

Given hallucination rates across 26 top models range from 22% to 94%, according to Stanford’s 2026 AI Index, released in mid-April, it’s an important problem to solve.

According to Moore, Elemental builds knowledge graphs from a customer’s own data, figuring out entities, relationships, and time and location attributes, as well as reporting the original sources of information.

For example, he said, “There are going to be 500 ships going through [the Strait of Hormuz, contested in the war with Iran] over the next week, and we know that some of them are going to have Iranian weapons in them. How do we figure out which ones are worth boarding to do a check on?”

With Elemental, the system can look at the history of the ships and their captains, their cargo manifests, even market and weather conditions, to figure out whether the ships are supposed to be where they are, or whether they’re acting in a suspicious way.

Elemental is not a replacement for an LLM, he explained, but something that helps companies improve the performance of the LLMs they’re already using. While it uses an LLM to help build the knowledge graph, once the graph is built, the process of using it involves traditional coding.

Read full article