Mphora’s VIVID platform lets product teams evaluate conversational AI by simulating virtual users built on over 1 million empirical personality profiles and the Big Five model. Its Persona Engine creates multi‑modal agents with memory and emotional state, while Compass and Shield provide longitudinal UX scoring and 5 000+ OWASP‑aligned security scenarios with automated compliance reporting.
Funding
Funding not disclosed
Founders
Product
Problem
Evaluating AI-driven conversational agents is difficult because user interactions are non‑deterministic, multi‑turn, and influenced by personality, context, and modality. Traditional testing methods that rely on static prompts or click‑path scripts cannot capture how real users experience, trust, or exploit such systems, leading to missed usability issues and security vulnerabilities.
Solution
Mphora’s VIVID platform provides a persona‑based simulation engine that creates virtual users grounded in over 1 million empirical personality profiles and the Big Five model. These agents interact with AI products across text, audio, video, and screenshots, retaining memory and emotional state over multiple sessions. The platform offers two evaluation products: Compass, which measures user experience through longitudinal, multi‑turn probing and generates per‑turn issue logs, and Shield, which runs more than 5 000 security and safety scenarios covering all 20 OWASP LLM categories with compliance mapping to EU AI Act, NIST AI RMF, and Korea AI Basic Act. By running persona panels in batch or swarm mode, developers obtain both quantitative reliability scores and a continuous stream of novel issue discoveries, enabling iterative improvement of AI agents before release.
Target Audience
Primary customers are product teams building conversational AI, LLM‑based services, and voice assistants who need rigorous UX and security evaluation, as well as compliance officers responsible for AI safety standards.
Features
- Over 5 500 virtual personas representing 25 cultures, each configurable for traits, linguistic style, and domain expertise
- Multi‑session memory with emotional state tracking (trust, frustration, fatigue) to simulate realistic user journeys
- Multi‑modal perception allowing agents to process screenshots, audio, and video alongside text
- Turing‑validated agents whose human‑agent similarity scores are statistically indistinguishable from human‑human interactions
- Compass panel scoring with 32‑judge panels that discover 4.9× more issues than single evaluators and provide per‑turn diary entries
- Shield’s 5 000+ OWASP‑aligned security and safety test scenarios, including batch and parallel swarm execution
- Automated compliance reporting mapping findings to EU AI Act, NIST AI RMF, and Korea AI Basic Act standards