"Owen Sakawa, the founder of Elloe AI, wants his platform to be the "immune system for AI" and the "antivirus for any AI agent." The idea, Sakawa said in an interview days before the TechCrunch Disrupt conference, where Elloe AI is a Top 20 finalist in the Startup Battlefield competition, is to add a layer to companies' LLMs that checks for bias, hallucinations, errors, compliance issues, misinformation, and unsafe outputs."
""AI is evolving at a very fast pace, and it's moving this fast without guard rails, without safety nets, without mechanism to prevent it from ever going off the rails," Sakawa (pictured above) said. Elloe AI is an API or an SDK, a module that sits on top of an AI model's output layer, an "infrastructure on top of your LLM pipeline," as Sakawa explained. "And it sits there basically fact-checking every single response.""
"The startup's system has its own layers, or "anchors," as Sakawa put it. The first anchor fact-checks the LLMs response against verifiable sources. Then, the second anchor checks if the output violates any regulations, such as the U.S. health privacy law HIPAA, the European far-reaching data protection and privacy law GDPR, or if it exposes some Personal Private Information (PII). The last anchor is an audit trail that shows how all the previous decisions were made and allows regulators or anyone auditing the system "to analyze the train of thought for that model from where it made the decision the source of that decision, the confidence score of all those decisions," according to Sakawa."
Elloe AI places an API/SDK layer on top of company LLMs to inspect every model response for bias, hallucinations, errors, misinformation, and unsafe content. The system uses multiple anchors: a first anchor to fact-check outputs against verifiable sources; a second anchor to detect regulatory violations such as HIPAA or GDPR and exposure of PII; and a final anchor that creates an audit trail showing decision sources and confidence scores. The technology relies on machine learning techniques and includes humans in the loop to maintain and update compliance and safety mechanisms.
Read at TechCrunch
Unable to calculate read time
Collection
[
|
...
]