\nArtificial intelligence is everywhere these days, quietly shaping everything from the way we write emails to critical decisions in healthcare and finance. But as AI becomes more woven into daily life, its downsides\u2014like hallucinations (those plausible-sounding but flat-out wrong responses), persistent biases, and the looming threat of data leaks\u2014have started to feel a lot more personal and risky.\n<\/p>\n
\nThat\u2019s where Hirundo steps in with something genuinely different. The Tel Aviv-based startup just secured $8 million in seed funding to tackle these AI headaches, with backing led by Maverick Ventures Israel and a strong bench of other investors. Instead of endlessly tuning models or filtering bad outputs, Hirundo is pioneering something called \u201cmachine unlearning.\u201d\n<\/p>\n
\nMachine unlearning is what it sounds like\u2014teaching AI to forget things it shouldn\u2019t know or behaviors we don’t want it to repeat. Imagine giving your AI model a targeted amnesia for bad habits, sensitive info, or unwanted bias\u2014after it\u2019s already been trained and deployed. No need to start from scratch or undertake a lengthy retraining process. It\u2019s a little like neurosurgery: the platform identifies the exact parameters inside a model that are triggering problems, then plucks those out with surgical precision, keeping everything else running smoothly.\n<\/p>\n
\nThis matters most in places where AI mistakes could lead to more than confusion\u2014think legal briefs, healthcare advice, or financial recommendations. Hallucinations in those fields aren\u2019t just odd\u2014they could mean lawsuits or shattered trust. Hirundo\u2019s approach means organizations can address these risks directly inside the AI, rooting out the causes rather than just patching over the symptoms. Early pilots in industries like banking, health, and even defense are already seeing models that produce more reliable, less risky results.\n<\/p>\n
\nWhat\u2019s more, Hirundo\u2019s technology is built to scale. It recognizes mislabeled data and weird outliers automatically, traces the roots of odd behaviors, and lets teams clean up AI models\u2014live, and often in just one step. There\u2019s no disruption to current systems and workflows. It works across data types, supports both generative and non-generative models, and can be deployed however security-conscious businesses want: as a SaaS tool, in their own private cloud, or even in locked-down, air-gapped environments that never touch the public internet.\n<\/p>\n
\nBehind the scenes are founders blending academic muscle and hands-on tech know-how: Ben Luria, Michael Leybovich, and Professor Oded Shmueli. With deep backgrounds in computer science, data security, and large-scale AI, they\u2019re well-positioned to steer the conversation around AI trust and reliability.\n<\/p>\n
\nIt\u2019s not surprising, then, that investors are paying attention. \u201cHirundo is taking on one of AI\u2019s most urgent challenges\u2014making sure these systems don\u2019t just sound convincing but are founded on truth, not discrimination or dangerous data,\u201d said Maverick Ventures\u2019 Yaron Carni. With the tech world waking up to the fact that AI trust is non-negotiable, Hirundo\u2019s vision for post-training machine unlearning feels not just timely, but necessary.\n<\/p>\n
\nAs AI\u2019s footprint expands into ever more sensitive domains, it’s clear that making AI \u201cforget\u201d its mistakes may be just as important as teaching it new things. Hirundo\u2019s approach points to a future where AI models can be both powerful and dependable\u2014a crucial leap for everyone who relies on the technology.\n<\/p>\n