Citadel AI was founded in Tokyo in December 2020 by Hironori “Rick” Kobayashi, a former trading-house executive, and Kenny Song, a former product manager for machine learning infrastructure at Google Brain. The two met online, discovered they both wanted to start an AI company in Japan, and built the business around a problem Song had seen firsthand: AI models that work fine in testing quietly degrade once they’re actually running in production, and almost nobody has good tools to catch it happening. According to the University of Tokyo’s own profile of the company, they spent a year in UTokyo’s FoundX startup support program before raising their first funding.

What It Actually Does

Citadel AI sells two main products that cover an AI system’s life before and after it goes live. Citadel Lens runs before deployment and during development: it stress-tests a model against industry standards and generates reports on explainability, robustness, fairness, bias, and data quality, the kind of documentation increasingly required by regulations like the EU AI Act and standards like ISO 42001. Citadel Radar takes over once a model is live, watching it continuously for drift, anomalies, and the kind of fraudulent or unusual input data that can quietly poison a model’s outputs over time. Both tools are described as model-agnostic, meaning they work by testing inputs and outputs from the outside rather than requiring deep access to a model’s internals, which the company positions as important because that kind of internal access tends to break every time a model gets updated.

Who’s Actually Using It

The clearest evidence that this isn’t just an idea on a pitch deck is who’s adopted it. Suntory Holdings, the Japanese beverage and spirits group, uses Citadel AI’s tools to monitor machine learning systems including a model that predicts pallet recovery in its logistics operations, and has taken an equity stake in the company. BSI, the British Standards Institution, selected Citadel AI from 54 companies worldwide after a year-long technical evaluation process to help assess high-risk AI systems in medical devices, automobiles, and credit screening. DeepEyeVision, a healthcare AI company, uses Citadel Lens for the kind of strict reliability assessment medical AI requires, and NTT Data has partnered with Citadel AI to verify Citadel Radar’s performance using open financial data.

The Open-Source Side

Unlike a lot of enterprise AI governance vendors, Citadel AI gives away a real piece of its technology for free. LangCheck is a Python library, released under an open license on GitHub, that provides building blocks for evaluating LLM-generated text: checking fluency, toxicity, sentiment, factual consistency, and more, with support for turning those checks directly into unit tests for an LLM application. It’s positioned as the foundation underneath Citadel AI’s paid Lens product, and the company has since built LangCheck Studio, a free browser-based playground where anyone can try LLM-as-a-judge evaluation techniques without installing anything.

What It’s Not

It’s not a household name, and it’s not chasing one. Total funding sits around $4.6 million, modest next to Japan’s bigger AI exports like Sakana AI’s roughly $2.65 billion valuation. The team is small, in the range of 11 to 50 people according to LinkedIn, and pricing isn’t public; the company quotes each customer individually based on use case. Coverage so far is limited to the University of Tokyo’s own feature profile, Tokyo startup job boards, funding databases, and niche newsletters, nothing from Verge, TechCrunch, Engadget, Wired, or Yahoo.

Who It’s Actually Built For

My Take — Mr Wangdoo

What’s genuinely interesting about Citadel AI, reading through its blog history and customer list rather than just its homepage, is how unglamorous the actual work is. There’s no flashy “AI agent” branding here, just years of incremental releases: Model Cards, Data Cards, object detection support, jailbreak testing in Japanese specifically, each one solving a narrow documentation or testing gap rather than chasing a headline. That’s not the kind of story that gets picked up by outlets chasing the next big AI announcement, but it’s exactly the kind of infrastructure that becomes load-bearing once enough companies are running AI systems they can’t afford to get wrong.

The detail I keep coming back to is the BSI selection, one of 54 companies evaluated worldwide, for a process that took a full year. That’s a slow, unglamorous kind of validation, the opposite of a splashy funding headline, but it’s arguably a more meaningful signal about whether the underlying technology actually works than any press release would be.