Korean Fintech AI Company Registers Patent Turning LLM Outputs into Default Probabilities
A Korean fintech AI company announced a registered patent that converts LLM outputs in credit review into measurable default probabilities using token log-probabilities and a lightweight calibration layer.
A patent registration announced on November 10 by a South Korean fintech AI company describes a method for converting large language model outputs in credit review into measurable default probabilities. The patent was reported under the name "Upin" by Korea Credit News and "Affinit" by Maeil Business; both outlets name the same CEO, Lee Chul-won, and the company's official name requires further verification.
Under the method, token log-probabilities for approval decisions, covering approve, reject, or review outcomes, are extracted and converted into a rejection probability. The raw values are not used directly, because the model's logic reflects pretrained statistics. Instead, a separate lightweight calibration function is trained on historical debtor delinquency data to map linguistic confidence to default levels. When market conditions change, only the calibrator is retrained, not the underlying LLM. The output is a quantitative default score used for approval thresholds and pricing per segment.
The filing acknowledges dependence on the quality of historical delinquency data, lag during shocks outside historical patterns, and regulatory explainability demands. The Korea Capital Market Research Institute has recorded that 11.2 percent of AI-based innovative financial services through August involve credit rating, loan, or insurance review.