Model Transparency
Understanding how our AI works and what you can expect from it
Kairos Model Transparency
At Kairos, we believe that artificial intelligence should be useful, understandable, and accountable. This page provides a transparent look into how our model works, what it's been trained on, and what you should (and shouldn't) expect from it.
What the Model Does
Kairos is designed to assist veterinary radiologists by generating concise conclusions and actionable recommendations based on your written findings. It does not interpret images or make autonomous diagnostic decisions. You're still the expert — we're just here to help you move faster without compromising quality.
Model Overview
Base Model: GPT-4.1-mini (OpenAI), fine-tuned for veterinary radiology
Model Version: Kairos-v1.1
Training Size: 5,000 anonymized radiology reports curated from board-certified radiologists
Modality Breakdown:
- - Ultrasound: 35%
- - Radiographs: 25%
- - CT: 25%
- - MRI: 15%
Input → Output: You provide findings; the model suggests conclusions and recommendations in real-time
Early Performance Snapshot
We are currently validating the model and actively seeking external research partners for external validation studies.
Qualitative issues observed:
- - Underemphasizing findings (e.g., suggesting findings may be incidental when they are not)
- - Incorrect recommendations
We're conducting structured evaluation using semantic similarity scores, expert panel review, and real-world feedback from active users. Validation results will be published next year.
Known Limitations
As with any generative AI model, Kairos has limitations:
- - It can't "see" — it depends entirely on your findings input
- - It may occasionally produce vague or overly confident language
- - It may not handle highly complex or unusual species/modality combinations well
- - AI output does not substitute for medical judgment, board-certified review, or client-facing communication
We've built in easy editing and feedback tools to ensure you stay in control.
Continuous Learning
Every time you edit the model's output or leave feedback, you help make Kairos better. We collect anonymized usage data and feedback ratings (positive / neutral / negative) to track performance and guide future training.
We'll be rolling out quarterly updates starting in Q3 2025, with transparent version notes on what's improved.
Data Privacy and Use
- - All training data is fully anonymized
- - Your reports are used for model training
- - We are compliant with PIPEDA and relevant privacy frameworks
- - Security is managed by Supabase and OpenAI, with encrypted data storage and role-based access control
Version History
| Version | Release Date | Notes |
|---|---|---|
| v1.0 | May 2025 | Initial fine-tuned model release. Optimized for small animal ultrasound, radiographs, and cross-sectional imaging. Includes structured findings-to-report flow and feedback tracking. |
| v1.1 | November 2025 | Trained on 5,000 radiology reports with improved coverage across CT and MRI modalities. Enhanced model architecture using GPT-4.1-mini for better performance and efficiency. Includes templated reporting workflow for AI-driven template population. |
Built by Radiologists, for Radiologists
Kairos was created by a team of veterinary radiologists who understand the pressures of high caseloads, tight turnarounds, and documentation fatigue. We believe that AI can support — not replace — our expertise. If you ever have questions, feedback, or want to join our testing group, we'd love to hear from you.