Measure
Reliability should be measured
Confidence in an AI answer should come from evidence, not from how fluent the answer sounds.
About Kanseko
Making AI answers something you can verify, not take on faith.
Kanseko is a Texas-based AI research and development company working on language model reliability.
Our story
From distillation to reliability
Kanseko started in model distillation research: making language models smaller, faster and cheaper to run. Smaller models raised a sharper question. When can you trust what a model tells you?
Our research into model reliability showed that it can be measured answer by answer, rather than guessed. That is now the focus of our work, and distillation remains part of our ongoing research.
The goal is simple to state: help the people who rely on language models spend less time checking answers that are fine, and catch more of the ones that aren't.
Principles
How we approach the work
Honesty
Show the full curve
We report results across every operating point, not a single cherry-picked number.
Practical
Research that leads somewhere
We focus on methods that can make language models cheaper, safer and more practical to rely on.
Founder
Fernando Canseco
Fernando Canseco is the founder of Kanseko. He holds a B.S. in Computer Science and an M.S. in Artificial Intelligence, with work spanning software engineering, machine learning, natural language processing, and computer vision.
Kanseko reflects his focus on model reliability, evaluation methodology, and efficient language model systems.
Contact
Interested in our work?
For more information about Kanseko and our research, email us.
contact@kanseko.com