Start with a question
Define the task, conditions, comparison, and acceptance criteria before interpreting a result.
HephaKnot is an applied AI research lab focused on evaluation and reliability.
We study how models behave under changing conditions, how the quality of data shapes their performance, and how evaluation methods can make their limits visible. We also develop Knotest, a platform for putting these methods into practice.
Our approach
Keep the evidence in view
Present findings alongside the method, scope, and supporting records. State where the available evidence stops.
Build useful instruments
Develop tools that connect data checks, model evaluations, and risk records in a repeatable process.
Research directions
We work in three directions: where models break, what the data misses, and what evidence a decision needs.
Model robustness
When do models become unreliable?
Investigate how noise, unfamiliar inputs, and changing conditions expose model failures, and how targeted training data can improve reliability.
Data quality
Which data should we collect next?
Explore how active learning identifies useful examples, helping teams improve model performance with fewer labelled samples.
Evaluation & assurance
What evidence supports trust in a model?
Examine how interpretability, privacy protection, and evaluation results can support decisions about using AI in practice.