How Hyderabadi Founder Convinced Mark Cuban to Solve Artificial Intelligence Hallucination

Months ago, TARGET watched many Indian founders make their way into the world’s largest startup accelerator Y-combinator. Many have since secured the backing of prominent investors to scale their ideas.

One of these founders is Kirill GorlaA 23-year-old prodigy from Hyderabad who mastered programming at the age of 11. He has now created CTGT, an artificial intelligence startup that has raised funding from high-profile investors such as Mark Cuban and Mike Knoop (co-founder of Zapier). .

Gorla and Trevor Tuttle founded CTGT, whose name includes their initials. The company was selected to compete at Y-Combinator in the fall of 2024 and was also named one of TechCrunch Disrupt’s top 20 Battlefield startups. Its goal is to improve the efficiency and interpretability of AI, and its model is already being tested by Fortune 10 companies.

The startup claims to eliminate AI hallucinations and make machine learning nearly 10 times more efficient. “CTGT’s goal is to make AI more transparent and accessible without sacrificing performance,” Gorlla said in an exclusive interview with TARGET.

Its platform integrates with both on-premise, open-source and API-based models, allowing enterprise customers to train custom, high-performance AI systems that deploy 10 times faster than traditional models.

Co-founders Cyril Gorlla and Trevor Tuttle

CGT and hallucinations

AI hallucinations have been a huge problem for everyone, including big tech companies like Google. Perplexity’s answering system recently experienced problems due to hallucinations fake news and citing them under real publications. Dow Jones and the New York Post filed a copyright infringement lawsuit against the company.

Interestingly, Microsoft filed patent for a method that claims to reduce and ultimately eliminate hallucinations.

Gorlla believes that existing solutions for LLM interpretability involve training hundreds of other models to identify concepts in the model, and then modifying each concept individually.

Since hallucinations are a universal problem for AI models, many research papers tried to solve the problem. CTGT solves this problem using a new platform that bypasses traditional deep learning methods, which often require huge computational resources and careful tuning to remove inaccuracies.

“These models are often not suitable for the business needs of enterprises, especially in mission-critical applications such as finance and healthcare. There is no last-mile level of ensuring that these models are at the level of reliability and reliability in terms of hallucination and brand consistency – that’s what we provide,” Gorlla said.

However, the platform instead directly examines the internal structure of the models, allowing companies to “control” the behavior of AI without introducing additional layers of complexity.

Exceeds the benchmark

Gorlla emphasized that many existing methods are computationally inefficient, “with state-of-the-art LLM interpretability from one leading fundamental model provider requiring more computation than the underlying model itself,” making such methods unaffordable for most companies.

“By focusing on understanding the underlying mechanisms of learning, we are building models from the ground up that are both powerful and interpretable,” Gorla said.

CTGT’s AI models were evaluated on several metrics. In a test of 121 classification datasets, training the neural networks took five hours, while the CTGT method took only 40 minutes.

“When tested on large (>50,000 training samples) classification and regression datasets, our method achieves accuracy of ≥ current transformer models, tree models, and MLP models,” Gorlla said. He also mentioned that the CTGT method required a total of 3,600 computing hours, while the other methods required 20,000 hours to set up.

Gorlla said the next version of their learning algorithm will be 500 times faster than the current one.

What’s the plan?

With a $500K deal from YC, $125K from Character Labs and an undisclosed amount from Cuban, CTGT is using its funds to refine its tech stack, expand its research and expand its customer base.

“Mark believes that interpretability and black box AI is a huge problem. We talked to him about what we were doing and he immediately agreed with our vision of truly intelligent AI and that the current status quo of deep learning will not lead to real AI,” Gorlla said.

His passion for programming and construction began at a very young age. “Growing up as an Indian-origin immigrant in the United States, I learned to code and at age 11 successfully completed a programming course at the local college where my mother lived, amid periodic utility shutdowns,” he said.

While still in high school, he saw the potential of artificial intelligence to drive global change and realized the importance of ensuring everyone had access to resources to benefit from it.