San Francisco – June 27, 2025 – CTGT, which permits enterprises to deploy AI for high-risk use circumstances, introduced at present an improve to its platform designed to take away bias, hallucinations and different undesirable mannequin options from DeepSeek and different open supply AI fashions. In testing, CTGT was confirmed to allow DeepSeek to reply 96 % of delicate questions, in comparison with 32 % initially.
CTGT demonstrated its new platform upgrades for the primary time at VentureBeat Remodel 2025, which came about June 24-25 in San Francisco. CTGT was considered one of simply seven finalists for the VentureBeat Innovation Showcase, and was named the highest winner of the Showcase by judges on the conclusion of the occasion.
In latest months, CTGT has constructed infrastructure for a number one insurance coverage supplier to make sure that their AI accurately coded complicated claims correctly, resulting in the insurer avoiding over $1M+ in monetary publicity in only one week. CTGT can also be architecting the ship of GenAI electronic mail compliance with a serious financial institution by robotically updating the corporate’s compliance fashions in manufacturing, with out having to take them offline.
In response to McKinsey, $67.4 billion in world losses have been linked to AI hallucinations throughout industries in 2024. Deloitte information has proven that 47 % of enterprise AI customers admitted they made no less than one main enterprise determination primarily based on hallucinated output. The issue seems to be worsening, with newer variations of some fashions hallucinating extra than the older ones. The not too long ago launched ChatGPT 4.5 mannequin has a 30 % hallucination fee – its worst ever. DeepSeek R1’s hallucination fee is 14.9 %, and far of its misinformation is intentional, coded into the mannequin by the Chinese language authorities.
In a high-stakes surroundings, these fashions can’t be trusted.
CTGT launched in February 2025 with $7M in funding from Gradient (Google’s early-stage AI fund), Normal Catalyst, Y Combinator, Liquid 2, Deepwater and notable angels together with François Chollet (Google, creator of Keras), Michael Seibel (Y Combinator, co-founder Twitch) and Paul Graham (Y Combinator). Initially CTGT was centered on decreasing the compute wanted to coach and deploy AI fashions – one thing it did fairly properly, decreasing compute by an order of magnitude by eliminating the necessity for fine-tuning, immediate engineering and backpropagation.
“DeepSeek’s introduction of R1 drove many enterprises to succeed in out to us on the lookout for assist – they have been concerned about benefiting from DeepSeek’s low value and velocity however involved about bias,” mentioned Cyril Gorlla, co-founder and CEO of CTGT. “In testing, CTGT was in a position to determine the precise mannequin options inflicting bias and take away them, in order that DeepSeek R1 may operate with out bias. CTGT has since carried out the identical exams on different open supply fashions together with Llama, proving that it could actually determine and take away bias, hallucinations and different undesirable options in minutes, with none time-consuming and dear fine-tuning or retraining. This functionality is now out there to any enterprise utilizing the CTGT platform.”
“AI mannequin trustworthiness is a serious impediment to enterprises reaching return on their AI investments,” mentioned Gorlla. “Every week I’m fielding calls from Fortune 500 executives wanting recommendation on easy methods to remedy this drawback. After doing customized work on this space for the previous few months, we’ve now productized our strategy into the CTGT platform and made it out there to anybody.”
