As an immigrant, Cyril Gorlla grew up learning to code and practicing like a man possessed.

"I aced my mother's community college programming course at 11, amidst periodically disconnected household utilities," he told TechCrunch.

In high school, he learned about AI, and was completely obsessed with the idea of training his own AI models that he started to break down his laptop for the internal cooling upgrade. Such tinkering landed him in an internship at Intel when he was on his second year in college to work on AI model optimization and interpretability.

At that time, he was witnessing the AI boom, when companies like OpenAI collected billions of dollars for its AI technology. Gorlla believed that AI could alter entire industries, but, above all, safety work in AI was lagging in the wake of new-fangled products.

I understood that there had to be this existence shift in the nature of how we think of training AI. Lack of certainty and trust is a huge barrier to uptake in industries such as healthcare or finance, where impact can be really quite dramatic.

So, with Trevor Tuttle whom he met as an undergrad, Gorlla quit his master's program to start company CTGT and help better deploy AI resources in orgs. Today, CTGT pitched their project at TechCrunch Disrupt 2024 through the Startup Battlefield round.
"My parents think I am in school," said the founder. "They'll be reading this is probably going to come as a shock to them.

It does this through collaboration with the firms so that it determines the hallucinatory model outputs as well as its biased version and then looks out for finding the root cause of the phenomenon.

Totally removing errors from the model is impossible. Yet, says Gorlla, CTGT's auditing technique can allow firms to empower themselves.

"We expose a model's internal understanding of concepts," he explained. "A model that tells a customer to put glue in the recipe is funny. What's not so funny when the response is about offering competitors as the recommended selection when the customer compares other products," he continued. "Outdated clinical studies for information for a patient, or credit decisions are made based on hallucinated info are not acceptable at all."

A recent survey from Cnvrg reported that enterprise adoption of AI apps claimed reliability was the top-of-mind concern. Risk management software provider Riskonnect found in a recent survey that more than half of the execs worry that employees are making decisions based on bad, inaccurate information from AI apps.

The idea of a separate platform to evaluate the decision-making of an AI model isn't new. Besides TruEra and Patronus AI, Google and Microsoft are also part of companies building tools that interpret model behavior.

However, says Gorlla, CTGT's techniques are more performant - partly because they don't rely on training "judge" AI to monitor in-production models.

Our mathematically-guaranteed interpretability differs from current state-of-the-art methods, which are inefficient and train hundreds of other models to gain insight on a model," he said. "As companies grow increasingly aware of compute costs, and enterprise AI transitions from demos to providing real value, our value is significant in providing companies the ability to rigorously test the safety of advanced AI without training additional models or using other models as a judge.".

To alleviate fears among potential customers about data leakages, CTGT presents an on-premises option, aside from the managed plan. It charges an annual fee for both.

"We don't have access to customers' data, giving them full control over how and where it's used," Gorlla said.

Graduating from the Character Labs accelerator, CTGT has investors like former GV partners Jake Knapp and John Zeratsky, co-founder of Character VC; Mark Cuban; and Mike Knoop, co-founder of Zapier.

AI that can't explain its reasoning is not intelligent enough for many areas where complex rules and requirements apply, Cuban said in a statement. I invested in CTGT because it is solving this problem. More importantly, we are seeing results in our own use of AI.

And — even though still very early-stage — CTGT has a number of customers, including three unidentified Fortune 10 brands. One of them, said Gorlla, worked with CTGT to remove bias from its facial recognition algorithm.
 
"We found bias in the model that focused too much on hair and clothing to make its predictions," he said. "Our platform provided the practitioners immediate insights without the guesswork and wasted time of traditional interpretability methods."

CTGT will, for the next few months, focus on building out its engineering team (which currently is only Gorlla and Tuttle) as well as optimizing its platform.

Should CTGT hit a toehold in the emerging market for AI interpretability, it's going to be very very lucrative. Analytics firm Markets and Markets forecasts that "explainable AI" as a sector may reach $16.2 billion by 2028.

"Model size is far outpacing Moore's Law and the advances in AI training chips," Gorlla said. "This means that we need to focus on foundational understanding of AI — to cope with both the inefficiency and increasingly complex nature of model decisions."