Enterprise leaders have joined a reliable program for nearly two decades. VB transform brings people to develop real enterprise AI strategies. Get more information
Based on San Francisco ctgtA startup that focuses on making AI more reliable through the feature level model customization, won the best presentation style award VB Transform 2025 In San Francisco. Founded by the 23 -year -old Serial Gorla, the company shows how its technology helps businesses overcome the barriers to AI Trust by directly modifying the features of the model rather than using traditional fine toning or instant engineering methods.
During his presentation, Gorla highlighted many businesses “Ai Dome Loop”: 54 % of the business deleted as the AI ​​as its highest tech threat, while McKenny reports that 44 % of the organizations have faced negative results from the AI’s implementation.
“A large part of this conference has been about Ai Dome Loop,” Gorla explained during his presentation. “Unfortunately, many of them (AI investment) do not go out. J&J now canceled Hundreds of AI pilots because they did not really supply ROI because of any basic confidence in these systems.
AI
The CTGT’s approach represents an important departure from the traditional AI custom technique. The company was founded at the University of California San Diego during a respected chair at Gorla.
In 2023, Gorla Published a dissertation Explaining a method of diagnosis and training of AI models at the International Conference on Learning Depression (ICLR), which was 500 times faster than the current point of view, while “three Nine” (99.9 %) obtain accuracy.
Instead of relying on the Bruit Force scaling or traditional deep learning methods, the CTGT has developed it, which it calls “a completely new AI stack” that basically reaffirms how the nerve networks learn. The company’s innovation focuses on understanding and interfering at the level of AI models.
The company’s approach is primarily different from the standard interpretation solution that relys on the secondary AI system for monitoring. Instead, CTGT offers a mathematical explanatory explanatory abilities that eliminate the need for additional models, which significantly reduce the computational requirements in the process.
This technology works by identifying specific variables (neuron or instructions in the feature space) that runs such as censorship or deception, then modifies these variables without changing the weight of the model. This approach allows companies to customize model behavior on the fly without taking offline system offline system.
Real -world requests
During his transform presentation, Gorla performed two enterprise applications already posted at the Fortune 20 financial institution.
An email compliance workflow that trains the model to understand the company’s acceptable content, which allows analysts to check their emails against real -time compliance standards. This system highlights potentially harassed content and provides specific explanations.
A brand alignment device that helps marketers to produce a copy according to brand values. This system can suggest personal advice on why certain phrases work well for a particular brand, and how to improve the content that does not align.
“If a company has 900 use issues, they will no longer have to fix the 900 models,” Gorla explained. “We’re a model-weed, so they can just plug us.”
A real -world example of CTGT technology in action was to work with it DPSEC modelWhere it successfully identified and edited the features responsible for censorship behavior. By separating and adjusting these specific activities patterns, CTGT managed to get 100 % response rate on sensitive questions without reducing the model’s performance on neutral tasks such as reasoning, mathematics and coding.
Photos: VB Transfer 2025 CTGT presentation
ROI demonstrated
It seems that the CTGT technology seems to be providing the measurement results. During the question and answer session, Gorla noted that in the first week of deployment, “one of the leading AI -driven insurances, we saved them $ 5 million.”
Another initial customer, Abrada Financial, has used CTGT to improve the accuracy of the facts of customer service chat boats. “Earlier, deception and other errors in response to a chat boot, Li Ibrada, said,” Earlier, a large amount of applications for direct support agents in response to a chatboat, as consumers tried to clarify the response. ” “CTGT has helped greatly improve the accuracy of the chatboat, most of them have eliminated the agent’s requests.”
In another case study, CTGT worked with an unknown Fortune 10 company to enhance the on -device AI capabilities in a computer limited environment. The company also helped a leading computer vision firm to achieve the performance of the 10 X -sharp model while maintaining comparison accuracy.
The company claims that its technology has reduced fraud by 80-90 % and can enable AI deployment with 99.9 % reliability, which is an important factor for businesses in regular industries like health care and finance.
From Hyderabad to Silicon Valley
Gorla’s journey is noteworthy itself. Born in Hyderabad, India, he Masked coding At the age of 11 and wasing a laptop in high school to squeeze more performance to train AI models. He came to study in the United States, the University of California, San Diego, where he got the chair fellowship.
The research there focused on understanding the basic method of learning the nerve network, which led to the ICLR paper and eventually CTGT. At the end of 2024, Gorla and co -founder Trever Tille, who specializes in the hyperclable ML Systems, was selected for the fall of the Y Combinator for 2024.
Startup has attracted significant investors outside its institutional supporters, including Mark Cuba and other prominent technology leaders who are attracted to its vision to make AI more efficient and reliable.
Fund and future
Gorla and Telt, CTGT established in the middle of 2024 7.2 million dollars collected In February 2025, Google’s initial stage led by the AI ​​Fund in an unmarried seed round. Other investors include General Catalist, Wii Combinator, Liquid 2, Deep Water, and Notable Angels such as Francois Chollet (Creasis Creator), Michael Sebell (Wii Combinator, Taving co -founder), and Paul Graham (Wii Combinator).
“The launch of the CTGT is timely because the industry struggles AI in the current boundaries of computing boundaries,” said Dariyan Shirazi, Managing partner of Milan. “CTGT removes the limits, which enables companies to rapidly measure their AI’s deployment and run advanced AI models on devices like smartphones. This technology is important for the success of AI deployment at high stake in large businesses.”
With the progress in Moore’s law and AI training chips to the AI ​​model size, CTGT aims to focus on a more fundamental understanding of AI, which can deal with both ineligible and rapidly complex model decisions. The company plans to use its seeds to expand its engineering team and improve its platform.
Each finalist presented the 600 industry decision makers and received feedback from a panel of sales force ventures, Menlo Ventures, and Venture Capital Judges.
Read about other winners Katio and Solo.u. Were other finalists ComoFor, for, for,. Superduper.ioFor, for, for,. Comedy And qdrant.
Editor’s Note: As thanks to our readers, we have opened the initial registration of birds for VB Transform 2026-only $ 200. From here, the AI’s desire is from operational reality, and you want to stay in the room. Keep your place safe now.
