Hirundo

What Hirundo does

Hirundo offers the first Machine Unlearning Platform designed to remove unwanted data and behaviors from AI models. It addresses issues like jailbreak vulnerabilities, hallucinations, bias, toxic outputs, and memorized PII by fixing risks at the model level. The platform allows for hardening models before deployment, fixing production issues without downtime, and continuously improving model behavior. Hirundo detects risky knowledge and behavior using built-in evaluations and red-team testing, then surgically targets and modifies specific parameters responsible for unwanted behavior, generating a fixed model in hours. It supports both LLMs and non-generative models, offering automated QA for vision, radar, timeseries, STT, and NLP datasets and models.

Key capabilities

• Machine Unlearning
• Detect risky knowledge and behavior
• Surgically target unwanted behavior
• Fix models in hours
• 85% reduction in jailbreaks
• 70% reduction in biases
• 100% PII removal
• LLM Unlearning
• Behavior Unlearning
• Data Unlearning
• Data QA for nongenerative AI
• Automated QA for vision, radar, timeseries, STT and NLP datasets and models
• Seamless integration with AI stack

Use cases

• Unblock launches in hours
• Harden models before deployment
• Fix model risks at the core
• Continuously harden in production
• Remediate issues in the model itself
• Enhance and customize model behavior
• Strengthen security
• Improve resilience
• Eliminate mission-critical risks
• Close SR 11-7 Compliance Gap

AI approach

Hirundo provides a Machine Unlearning Platform that allows users to detect and fix risks in AI models by surgically modifying specific parameters responsible for unwanted behavior, without retraining the entire model. This helps in reducing jailbreaks, biases, and removing PII from LLMs and other non-generative AI models like vision, radar, timeseries, STT, and NLP datasets.

Tech named: Machine Unlearning, LLM Unlearning, Behavior Unlearning, Data Unlearning, Data QA for nongenerative ai

What it says sets it apart

• First Machine Unlearning Platform
• Fixes risks at the model level, not just the perimeter
• Significantly faster than retraining (hours vs. months)
• No need to change workflows for integration
• Maintains model utility and performance while fixing issues
• Supports both LLMs and non-generative models

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.