Hirundo

Hirundo develops a machine unlearning platform that surgically removes risky knowledge and behaviors from trained AI models at the model level, without requiring retraining. The platform addresses critical AI safety and compliance challenges including jailbreak vulnerabilities, hallucinations, bias, toxic outputs, and memorized PII (personally identifiable information). Hirundo's three-step workflow detects risky knowledge and behavior through built-in evaluations and red-team testing, surgically targets it with machine unlearning that identifies and modifies the specific parameters responsible for unwanted behavior without harming model utility, and delivers a fixed model in hours rather than the months required for retraining. The platform delivers measurable results including 85% reduction in jailbreaks (verified on benchmarks like PurpleLlama), 70% reduction in biases (verified on BBQ benchmark), and 100% removal of fine-tuned PII with zero impact on other data or functionality. Hirundo serves three use cases: unblocking launches in hours not weeks (hardening models before deployment so late-stage issues do not derail launches), fixing model risks at the core (removing memorized PII, reducing jailbreaks, correcting bias at the model level rather than the perimeter), and continuously hardening in production (fixing production issues and improving model behavior without downtime). The platform also offers automated data QA for non-generative AI including vision, radar, timeseries, STT, and NLP datasets and models. Hirundo has been featured by Google DeepMind and has demonstrated increasing the security robustness of Liquid AI's LFM 2.5 by 92%. Founded in 2023 and headquartered in London, United Kingdom, Hirundo has raised $8 million in a Seed round backed by Maverick Ventures, SuperSeed, AIC, Tachles VC, AI.FUND, and Plug and Play Tech Center.