Hirundo, the first AI startup specializing in machine unlearning, has raised an $8 million seed round to combat one of AI’s most pressing challenges—how to make AI forget. Backed by Maverick Ventures Israel, with support from SuperSeed, Alpha Intelligence Capital, AI.FUND, Tachles VC, and Plug and Play Tech Center, Hirundo is pioneering tech that retroactively removes biased, malicious, or confidential data and hallucinations from trained AI models—without retraining and without degrading performance.
Why AI Needs to Forget
As generative AI shifts from experimental tools to mission-critical infrastructure, major issues like hallucinations, bias, and prompt vulnerabilities threaten enterprise trust. Studies show:
- Nearly 40% of AI-generated “facts” contain bias
- Over 50% of U.S. employees worry about AI inaccuracy
- Legal, reputational, and operational risks are increasing as AI deployment expands
Hirundo addresses this risk not before model training—but after. Unlike conventional tools that filter or mask issues, Hirundo performs precise behavioral neurosurgery on models, targeting the exact regions in the model’s billions of parameters where hallucinations or toxic outputs arise.
The Technology: AI That Can Forget
What Is Machine Unlearning?
Hirundo’s machine unlearning system works across both open-source (e.g. Llama, Mistral, Gemma) and soon gated models (e.g. ChatGPT, Claude), enabling:
- Removal of poisoned, biased, or confidential training data
- Reduction in hallucinations and prompt injections
- Preservation of core model performance—no retraining needed
Technical Impact
In early deployments, Hirundo has achieved:
- Up to 70% bias removal (tested on DeepSeek-R1)
- 55% hallucination reduction
- 85% decrease in prompt injection success (tested on Llama models)
This allows enterprise AI systems to be more secure, fair, and compliant—without expensive re-engineering or retraining.
Use Cases & Adoption
Hirundo is fast becoming the first-choice solution for high-stakes AI applications in:
- Finance (e.g. fraud detection, algorithmic trading)
- Healthcare (e.g. diagnostics, patient-facing chatbots)
- Defense and Aerospace
- Government and Public Sector
- Consumer Platforms
Its platform supports both generative models (like LLMs) and non-generative systems (e.g. computer vision, radar, LiDAR, NLP).
Industry Reactions
Ben Luria, CEO & Co-Founder of Hirundo, said:
“Hirundo’s solution works like AI neurosurgery—pinpointing where toxic knowledge is encoded and removing it without damaging model performance. This funding will accelerate our mission to enable AI systems that are fair, safe, and truly enterprise-ready.”
Yaron Carni, Founder, Maverick Ventures Israel, commented:
“Without removing hallucinations or biased intelligence from AI, we encourage mistrust and risk real-world consequences. Hirundo’s unlearning approach is transformative.”
Mads Jensen, Managing Partner, SuperSeed, added:
“We back AI companies transforming industries—but transformation must be built on trust. Hirundo directly solves AI’s most critical reliability gap.”
Founding Team
Hirundo was founded in 2023 by an elite trio:
- Prof. Oded Shmueli: Former EVP of Technion and leading authority on data theory
- Ben Luria: Serial entrepreneur, Rhodes Scholar
- Michael Leybovich: Expert in data lineage and enterprise AI infrastructure
As AI becomes ubiquitous, hallucinations, bias, and data toxicity are no longer tolerable side effects—they are liabilities. Hirundo’s machine unlearning technology provides the missing piece of AI reliability and governance. With $8M in fresh capital and a world-class founding team, Hirundo is set to become the AI safety standard for enterprises and governments worldwide.
Explore how TechEdgeAI unlocks transformative insights, streamlines operations, and drives unparalleled innovation for your business. Visit our website or connect with us to discover the future of intelligent solutions.
