trismik
Research led
38 +
years of combined research experience
200 +
published research papers including ACL, EMNLP, NEURIPS, and AAAI
12,000 +
author citations
We believe that LLMs should be useful and safe for everyone. As a Cambridge University spin-out we have a track record of contributing NLP resources and methods to the R&D community. At Trismik we focus on efficient evaluations of real-world tasks using a mix of human- and LLM-judgements to deliver automated testing in a wide range of settings.
Founding team
Rebekka Mikkola (CEO)
Nigel Collier (CSO)
FAQs
What's your back story?
Trismik is an early stage spinout company, founded by a group of technology geeks with roots in Cambridge University, Salesforce and Amazon. Backed by 38+ years of research experience we're dedicated to building tools that will lead to a future where AI can contribute to human well-being.
What sets Trismik apart?
Our adaptive testing system uses real-time interactions with your model to dynamically adjust the difficulty of questions based on previous responses. This gives you a more comprehensive view of how your model will perform in real-world conditions.
How does Trismik deliver up to 100x faster evaluations?
Our tests are delivered in 60 items or less to obtain an ability score, compared to an average basket of benchmarks with 17581 items (MMLU, Math, AI reasoning, GSM8K and Winogrande). This results in >290X fewer test items. For obtaining model comparisons we recommend taking an average of 10 or more test run results. Actual test times will vary according to many factors including prompt and output length, tokenization method, model size, memory bandwidth, batching method, as well as network latency.