Description

This talk explores two critical AI safety challenges—hallucination and prompt injection—through the lens of multi-agent architectures and open-source standardization. A multi-agent pipeline for hallucination mitigation is presented, based on research (Gosmar et al., 2025, https://arxiv.org/abs/2501.13946), which employs iterative review loops, automated fact-checking mechanisms, and explicit disclaimer insertion to refine outputs and enhance reliability. A set of novel Key Performance Indicators (KPIs) guides the process, ensuring that speculative content is appropriately grounded in verifiable sources. Takeways: This presentation is based an original research paper recently published by me, that explore critical AI safety challenges—hallucination mitigation—through the lens of multi-agent architectures and open-source standardization. The paper, “”Mitigating AI Hallucinations via Multi-Agent Fact-Checking and Review Loops””, introduces a structured pipeline for enhancing AI reliability by employing iterative review loops, automated fact-checking, and explicit disclaimer insertion. (Gosmar et al., 2025, https://arxiv.org/abs/2501.13946) What makes this approach unique is the application of standardized inter-agent communication, leveraging the Vendor-Independent OFP framework to enhance AI governance, security best practices, and trustworthiness in generative AI systems. This structured approach ensures greater interoperability, resilience, and transparency in AI-driven solutions.