YRIKKA: Automated AI Safety and Reliability for High-Stakes Applications
Our investment in YRIKKA
Today we’re thrilled to announce our partnership with Kia Khezeli, John Kalantari, and the entire YRIKKA team on their mission to bring confidence to deploying computer vision models into production in high-stakes scenarios.
Kia Khezeli, Ph.D., John Kalantari, Ph.D., co-founders of YRIKKA, and Arpan on a visit to the YRIKKA office in NYC
Why Test and Evaluation for Computer Vision AI Models Matters
While the dramatic breakthroughs in large language models have been the ones making headlines over the last couple of years, computer vision models have been advancing rapidly in parallel. Generative AI vision models, and newer breakthroughs that apply the transformer architecture to vision tasks like image classification and object detection are just two such examples. And the impact of computer vision - or “Visual AI” models are by nature “real world” in nature - verticals like defense, healthcare, automotive, and agriculture are all huge markets that benefit from visual AI breakthroughs.
But deploying visual AI in real-world, mission-critical settings remains hard – not because the models aren’t good enough, but because no one’s quite sure how good they are in the contexts that matter most. In high-stakes scenarios like defense and healthcare, you have to be sure you’ll get it right, or you just won’t push to production.
As a result, while testing and evaluation tools have emerged in the context of language models, the analogous frameworks in visual AI are still lacking. The current state of the art here consists of brute force - throwing expensive ML talent at the problem - or wiring up generic MLOps tooling that isn’t tailored to visual AI.
YRIKKA replaces these approaches with something faster, more focused, and more holistic.
YRIKKA’s AI Evaluation and “Red Teaming” Capabilities
By integrating synthetic data generation, adversarial red teaming, and fine-tuning capabilities, YRIKKA significantly reduces the time required to validate AI models – from months to minutes. YRIKKA’s approach ensures that breakthrough visual AI models are safe, reliable, and as performant in real-world scenarios as they are in benchmarks.
YRIKKA’s core idea is simple but powerful: "contextual red-teaming" for visual AI models. YRIKKA automates the painstaking work of evaluating visual AI models under real-world constraints. Instead of testing models in broad, abstract settings, YRIKKA evaluates how a model performs in the exact context it’s meant to be deployed in. That includes everything from identifying edge cases to synthetically generating images that mirror the specific challenges a model might face in production. The result: a platform that helps teams pick the right model, test it thoroughly, identify failure modes, and fine-tune it—all in a fraction of the time it currently takes.
YRIKKA’s platform enables AI developers to automatically “red-team” a visual AI model to uncover vulnerabilities and biases that would impact reliability and accuracy
This allows AI teams to move beyond general-purpose benchmarks and into tailored, confident deployment. In other words, YRIKKA helps them go from “this model seems good” to “this model is good for us.”
Deeply Technical Founders with Commercial Aptitude
YRIKKA was co-founded by Kia Khezeli and John Kalantari, both of whom bring deep technical experience and domain expertise to this problem space. Kia earned his Ph.D. in Electrical and Computer Engineering from Cornell, focusing on safety in AI systems, and later joined Mayo Clinic as a Principal ML Scientist in their Biomedical Artificial General Intelligence Lab. John holds a Ph.D. in Biomedical Engineering and has worked as an AI consultant at NASA and across the Mayo Clinic’s AI teams.
Together, they’ve designed YRIKKA not just as a research tool, but as a product that can slot into real-world workflows. From our early conversations with the team, it’s clear they’re not trying to build an academic curiosity—they’re going after a commercial opportunity with real urgency and clarity.
Kia and John’s backgrounds reflect the product’s focus on high-stakes verticals such as defense, healthcare, autonomous systems in manufacturing, and agriculture. These are sectors where failure is expensive, regulation is meaningful, and trust is everything. As a result, YRIKKA is well positioned to help accelerate the deployment of all of the visual AI innovations that are emerging, and thus have a significant impact on the broader adoption of AI into society and the global economy.
Let’s Go!
We at Garuda Ventures are thrilled to have co-led YRIKKA’s pre-seed round alongside our friends Pascal and Daniel at Focal.
The company’s “APEX” (Adversarial Probing and Evaluation of Exploits) API is publicly available as of today. Reach out to us if you’d like to be connected with the team!
Garuda Ventures is a B2B focused pre-seed fund, most often leading a founder's initial round of financing as their first true believer. The firm aims to be a founder's first call as they navigate their way from a deck and a dream to building a category-defining company. To learn more, visit https://www.garuda.vc