Hirundo – Machine experts learned $8 million The Seed Funding Round, led by Maverick Ventures Israel, has participated from SuperSeed, Alpha Intelligence Capital, Takuls VC, AI. Fund and Plug and Play Tech Center. The company is developing an industry-first solution to “forget” your malicious, sensitive data from trained AI models, biases, and hallucinations. This is a problem with AI models that generate misleading or inaccurate information that appears factual. Hirundo accomplishes this without retraining the model from scratch and without slowing down the performance of the model, thereby de-employing AI.
Also Read: Is Lora the backbone of a distributed AI network?
As generator AI evolves from experiments to enterprise infrastructure, serious flaws begin to appear, and once models are trained, tweaked, or deployed, it is almost impossible to correct that mistake. Researchers found that nearly 40% of the “facts” used by AI contain bias, but half of US employees cite inaccuracy as a major concern associated with generated AI. As AI becomes wider and more powerful, more hallucinations and biases become apparent, leading to worsening hallucinations than ever before, leading to legal, reputational and operational risks for the companies that rely on them. Hirundo is tackling the AI flaws from a completely unique starting position after the model has already been trained. By retrospectively removing unwanted behaviors and inaccuracies from AI models, there is no need to interfere with the functioning of the model or retrain it. This is a time-consuming multi-million dollar movement, and Hirundo will improve AI performance and reliability, ensuring that businesses trust and deploy AI.
There are various techniques for fixing AI model problems and spans the entire AI development lifecycle, but these do not prevent hallucinations and biases from leaking into the final AI project. Traditional solutions like Guardrails and fine tuning are insufficient. The root cause of the model itself remains untouched, but it simply masks and filters bad data or behavior. Hirundo sees prison escape from an action perspective. Using its machine's unlearning approach, the company finds instructions for models that are prone to hostile manipulation and removes them. This is intended as a future AI model for attacks.
“The broader adoption of AI is limited by hallucinations and unwanted behaviors that are too risky to deploy models to enterprise-level applications. Hirundo allows you to instantly improve your models at the core and work towards a more equitable and accurate output. “Hirundo's solutions work like the form of the AI model 'neurosurgery', ensuring that hallucinations occur or toxic knowledge are encoded and accurately deleted, ensuring that data is deleted and the accuracy of the model is guaranteed.” ai. ”
Read again: Upgrade to smart meeting rooms with AI integration
“We're distorting the outcome and promoting distrust without removing the intelligence of hallucinations and bias from AI. For organizations and businesses, this means legal, financial and real-world outcomes. The majority of companies suffer from AI built with bias, but most companies cannot fully address it. To invest and support their journey before them.” Yaron Karnithe founder of Maverick Ventures Israel, leads this funding round.
Mads JensenSuperseed's managing partner commented. “We are investing in exceptional AI companies that transform the industry in industry, but this transformation is as powerful as the models themselves are trustworthy. Hirundo's technology is groundbreaking to address the issues of hallucination and bias that plague even today's most advanced models.
Hirundo can identify how a particular behavior presents in the model, utilize industry-accepted benchmarks, outline unwanted behavioral traits, and then move the model away from them. This approach works with both open source models (such as Llama, Mistral, Gemma) and soon gate models (such as ChatGpt, Claude). When deployed, Hirundo's solution removed bias by up to 70%, resulting in a reduction of up to 55% in hallucinations and an 85% in successful rapid injections, as demonstrated in a study with DeepSeek-R1.
Hirundo, the first choice solution for mission-critical AI deployment in industries such as finance, healthcare and other consumer industries, as well as high-risk enterprise and defense applications, already pilots a variety of multinational companies and government agencies. Hirundo's platform supports both generative models (such as large-scale language models) and non-generated systems used in areas such as computer vision, radar, LIDAR, and NLP.
Founded in 2023 by Professor Emeritus oded shmueliformer dean of computer science and EVP. Israel's Technion together Ben Luriaserial entrepreneurs and Rhodes scholars, and Michael LeibovichHirundo, an expert in the field of data lineage, is the world's first startup to pioneer the concept of “making AI forget” and provide unlearned solutions.
[To share your insights with us as part of editorial or sponsored content, please write to psen@itechseries.com]
