Citadel Lens and Citadel Radar AI Trust Layer

Citadel AI Inc. is a Tokyo-based startup incorporated in December 2020 by CEO Hironori Kobayashi and CTO Kenny Song. It provides a technical trust layer that helps organizations test, monitor and govern machine learning and generative AI systems from proof of concept to production, and has worked with NEDO, AIST and Japan's Cabinet Office on AI safety evaluation.

Citadel Lens evaluates LLM, RAG and computer vision systems with automated metrics for hallucination, relevance, toxicity and jailbreaks, an 11-category safety dataset, automated red teaming, no-code custom metrics and human-in-the-loop review, then produces technical and compliance reports aligned with Japan's AI Guidelines for Business and the EU AI Act. Lens Governance adds AI inventory and pre-deployment review, while Citadel Radar acts as a real-time AI firewall that blocks jailbreaks, PII and data exfiltration. Users include Tokio Marine, ONO Pharmaceutical and BSI.

• Automated LLM metrics: scores hallucination, relevance, toxicity and jailbreak resistance in multiple languages.
• Safety dataset and red teaming: 11 risk categories and adversarial prompts built from jailbreak templates.
• Custom metrics: prompt-defined evaluation criteria without programming.
• ML model testing: robustness checks, automatic bias finder and prediction explanation for vision and tabular models.
• Lens Governance: AI inventory, risk assessment and pre-deployment review workflows.
• Citadel Radar: runtime AI firewall and agent monitoring that flags or blocks unsafe inputs and actions.

Vietnamese banks and insurers can test customer-service chatbots for hallucination and jailbreaks and keep audit-ready reports. Japanese-invested manufacturers in Vietnam can validate visual inspection models for robustness and bias. Certification and audit firms can use the platform to assess clients' AI systems.

Risk and compliance teams get standardized, ISO-referenced test reports instead of ad hoc checks before launching GenAI services. Automated metrics and red teaming cut manual review effort while human annotation keeps scores trustworthy. A single loop from pre-deployment testing to runtime firewall reduces the chance of jailbreaks or data leaks reaching customers. SaaS, private cloud and on-premise options suit regulated sectors.

Regulators in Japan, the EU and Southeast Asia are moving from voluntary AI principles to testable requirements, and certification bodies now need tools to audit high-risk AI. Banks and insurers in Vietnam and the region rolling out GenAI assistants face similar pressure to show evidence of testing and runtime controls.

Citadel AI Inc. was incorporated on 10 December 2020 in Bunkyo, Tokyo, and is not affiliated with Citadel Securities. It has run research with NEDO and AIST and a Cabinet Office project on AI evaluation environments. BSI adopted Lens for EU AI Act conformity assessments, and CHAI lists it as an assurance resource provider. The site shows HIPAA, GDPR and ISO 27001 badges.