How One Founder Is Fixing AI’s Biggest Threat: The Black Box Problem

What we talked about.

This episode of the Sunny Ray Show features Alex, founder of an AI safety and security startup built around detection and response for AI systems, from simple chatbots to complex agentic workflows. Alex traces his path from coding BASIC on a Commodore 64 as a child, to working at Austria's first internet service provider, to Cisco networking, cybersecurity leadership, and eventually AI explainability research at the University of Houston. He explains why using large language models to police other large language models is risky, arguing instead for deterministic, explainable machine learning paired with generative context. The conversation covers agentic AI observability, the compressed regulatory timeline compared to the internet and social media, why AI models remain black boxes despite being built from transistors and code, and how the company deploys its ten million dollar seed round toward go to market automation. Alex also reflects on why he still teaches as an adjunct professor and how student talent shaped his company's early research team. It is a grounded look at the technical and human challenges behind building trustworthy AI.

A cybersecurity veteran explains why AI's black box problem demands explainable systems, not more black boxes watching black boxes.

The questions, and the answers.

Before we dive in, can you share what your company does in a sentence or two?

We do detection and response for AI events. An AI event can be as simple as a chatbot, a prompt and response between a user and a large language model, or as complex as agentic workflows where an agent talks to memory, to other agents, pulls content from the internet, or executes a tool. Our job is to actively monitor these events and detect safety and security issues, whether that's disclosure of PII, PHI, PCI data, or active attacks against the model.

What was it about the Commodore 64 that grabbed you as a kid?

My dad worked for IBM and I remember, at maybe five or six, seeing his massive terminal in the basement. The Commodore 64 was accessible, you could play games, but it also opened the door to writing your own software. I wrote simple programs to time running races at school. It was the magic of discovering new things daily and not knowing where it would end that fascinated me.

What did working on mainframe automation teach you that still sticks today?

I learned coding wasn't my future. I spent a lot of time writing code, then the company hired a mathematician to optimize it because it wasn't fast enough. That was the moment I realized I could never be the best at coding since I lacked some of the classical math foundations. It pushed me toward network tech instead, which eventually led me to work at an early internet service provider.

What made you decide networking was more exciting than development, and what did Cisco teach you?

It was the new shiny thing, honestly. After three or four years in coding, the internet and routers came up and people needed engineers who could program Cisco devices. We were asked to do pen testing for a local bank at 22, and it was a tremendous amount of fun. That sense of unlimited possibility is what pulled me deeper into network tech and eventually into freelancing for Cisco.

What specific moment made you think you needed to build something new in AI safety?

I'd always been involved in AI within cybersecurity, seeing unsupervised learning move into the endpoint years ago. As I got deeper into deep learning and neural networks, I became fascinated by how we keep control as this accelerates if we can't understand how these models operate. That led me to raise a million dollars and start a research project on AI explainability in a lab in Houston, right before GPT launched and everything pivoted to GenAI.

You've talked about the problem of using one black box to monitor another black box. When did that insight crystallize?

Step back from AI for a second. If you discover a highly advanced technology, would you use that same technology to protect yourself from it? Intuitively, that doesn't make sense. I wouldn't call it a trap so much as the space being so new that people don't distinguish between an LLM based assessment and a machine learning based one. We use good old fashioned machine learning so we can give full explainability, even without the verbosity of an LLM response.

These are just transistors, ones and zeros. How did we end up with black boxes even top researchers can't explain?

Large language models are transformers, essentially neural networks with many layers of neurons performing highly specialized functions, gate functions, binary functions. Information trickles through so many layers and neurons that you can't reverse engineer the output back to what happened at each step. There is research into modular architectures and world models trying to build explainability in from scratch, so yes, I think we can get there, but right now it's the speed and complexity that outpaces our ability to interpret it.

You raised a ten million dollar seed round from SYN Ventures. What's the plan for deploying that capital?

The majority goes into go to market. We spent over two and a half years building research and applied data science teams, and now it's about translating those capabilities into customer facing products. We built a full RevOps platform using agents for account research and personalized outreach, which lets a startup like us operate at a fraction of the headcount that would have been needed for the same output ten years ago.