Clinical AI
at the frontier
We train hospital-native multimodal AI in a virtual clinic and deploy it as a dependable coworker within kaiko.w to assist healthcare professionals with administration, documentation, and report generation so they can focus on patient care.
Our beliefs
Environments matter
Future AIs will evolve and specialize according to the environments they operate in. We are training our models directly inside a realistic simulation of hospital IT systems: the kaiko.ai virtual clinic. This enables our agent to seamlessly operate in a hospital just like a healthcare professional would and execute long-running, complex tasks autonomously.
Systems over algorithms
Modern AI is about designing systems, not algorithms. Without a stable, scalable infrastructure capable of automating large parts of data processing and synthetic generation, evaluation, and training – algorithmic tweaks are futile. At kaiko.ai, we focus on building systems first, algorithms follow.
Safety and faithfulness
Designing safe and trustworthy AI for healthcare is non-negotiable. We are building extensive benchmarks reflecting real clinical workflows to measure safety, accuracy, and faithfulness. We also perform research into effective oversight mechanisms at inference time such as CoT monitoring.
Scale
The scaling hypothesis is true. Most of healthcare AI is stuck in the past, training small models for narrow use cases while the rest of the field has moved on. At kaiko.ai, we take this seriously and focus on building healthcare-specific AI at frontier scale and generality.
Our models
Model 1
Released
kaiko-ai/midnight-1
A vision foundation model for pathology based on DinoV2.
Publications
Training State-of-the-Art Pathology Foundation Models with Orders of Magnitude Less Data
The paper accompanying the release of our pathology vision model “Midnight”: SOTA performance can be reached with orders of magnitude less data than previously thought.
Evaluation Framework for Pathology Foundation Models
Our open-source framework: eva, for evaluating computational pathology foundation models (FMs).
Towards Training Large-Scale Pathology Foundation Models: from TCGA to Hospital Scale
Initial work towards our first pathology vision FM preceding and laying the groundwork for our most recent Midnight family of models.