Most content classifiers are built for a fixed list of categories. CriteriaBot was built to work for whatever category you define.
Graham Paye needed a content moderation tool that could catch fallacies, cognitive distortions, persuasiveness signals, factual correctness, and on-topic-ness - on top of the standard harmful content categories most services cover. Nothing on the market came close. The services that existed had narrow category lists, enterprise pricing that required a sales call, and no meaningful customization. So he built his own.
The result is CriteriaBot, a universal content classifier that works across any content type and any category or criteria a user defines. The core architecture runs a pool of small, open-source language models as a consensus system, then threads their individual votes through traditional machine learning approaches to draw a final conclusion. That layered design is what makes it generalizable: because the system is drawing consensus rather than pattern-matching to a fixed taxonomy, it handles novel categories with high accuracy out of the box - no retraining required.
Personalization That Kicks In Immediately
The harder problem in content moderation is judgment calls - categories where reasonable people disagree. CriteriaBot addresses this with a personalization layer that factors in user feedback at request time. When a user flags a verdict or provides examples of how they'd classify something, the adjustment is applied from the very next evaluation. There is no retraining cycle, no waiting. The system recognizes where its defaults might diverge from a specific user's standards and corrects in real time.
For queries that require current factual context or mathematical reasoning, CriteriaBot pulls from Wikipedia and Wolfram rather than relying on model training data alone. The underlying models are also multi-modal, making it straightforward to extend classification to different content types beyond text.
The use cases Paye sees beyond content moderation include brand voice enforcement and prompt injection detection - both categories where existing tools are either too rigid or too expensive to deploy at scale. CriteriaBot is available now at criteriabot.io.
