About Telluvian
We're a research-focused team dedicated to making AI systems trustworthy for high-stakes domains through mechanistic interpretability.
Our Story
Telluvian was founded on a simple observation: as AI becomes central to legal practice, the stakes of AI failures grow dramatically. A hallucinated case citation isn't just an error—it can result in court sanctions, malpractice claims, and damaged client relationships.
Traditional approaches to AI safety can't solve this problem. Prompt engineering and output filtering treat models as black boxes. We take a different approach: using mechanistic interpretability to understand what's actually happening inside AI models, we can detect hallucinations as they occur.
Expertise
Mechanistic Interpretability • AI Safety • Legal Technology
Focus
Hallucination Detection • Model Understanding • Trustworthy AI
Our Mission
Making AI trustworthy through understanding.
We believe the path to trustworthy AI runs through understanding how AI models actually work. Rather than treating them as black boxes to be monitored from the outside, we develop techniques to look inside and understand their computations.
Mechanistic interpretability is our core methodology. By training probing classifiers on model activations, we can identify the specific patterns that indicate when a model is confabulating or uncertain.
Our initial focus is on legal AI, where the consequences of hallucinations are severe and well-documented. But our approach applies broadly to any domain where AI trustworthiness matters.
