Citadel AI, the Japanese startup developing automated AI quality maintenance tools, announced on Monday that it has secured 100 million yen (about $900,000 US) in a seed round from UTokyo Innovation Platform (UTokyo IPC) and Anri. For the startup, this is the first funding from external investors. They launched Citadel Rader in beta in May, aiming to help companies protect themselves from AI-specific risks by automatically monitoring their AI systems, detecting, blocking, and visualizing anomalies.
Citadel AI was launched in December by CEO Hironori “Rick” Kobayashi and CTO Kenny Song. Prior to Citadel AI, Kobayashi served Loyalty Marketing as president, Mitsubishi Corporation (Americas) as SVP, and US-based meat processing firm Indiana Packers Corporation as CEO. Meanwhile, Song led the development of TensorFlow and AutoML as a product manager at Google Brain, the tech giant’s AI research and development unit.
Unlike traditional hardware-based software, AI systems are exposed to an ever-changing real-world environment that degrades their accuracy and quality day by day. It is important for businesses to maintain the quality of AI functions by automatically detecting anomalies before they are misrecognized and misjudged, resulting in business losses and compliance issues. Citadel Rader has an XAI (eXplainable Artificial Intelligence) function that automatically detects and blocks AI input and output anomalies and visualizes them in a form that humans can understand.
In the development stage, AI reads only clean data, but when it moves to actual operation, it receives a variety of data, including those with input errors. Basically, people think that computers will give correct answers, and even if they give wrong answers, it is difficult to point them out.
Since it is difficult for companies to allocate human resources to monitor the output of AI, our tool may help AI engineers who are usually busy with their daily work find the time to concentrate on their original work.
When a system integrator receives an order for an AI system, they will typically implement the system but not provide services to automate the operation and maintenance afterwards.
If the accuracy and quality of the data deteriorates, in the worst case scenario, it could lead to errors in sales forecasting, or in credit approval. For example, think FATF (Financial Action Task Force, the global organization working with money laundering regulators in various countries). A single node with poor security in determining a money laundering case could lead to the vulnerability of the entire global network, which could lead to the node not being allowed to join the organization.
He added that Citadel Rader is currently used by more than 10 companies on a trial basis and is in talks with more than 100 companies as potential users. The company plans to use the funds to expand its engineering team for the product’s official launch which is scheduled next spring.