Perpetual, Inc.
Perpetual develops AI systems that continuously learn and adapt after deployment by treating four established techniques—self-rewarding, RLHF, distillation, and RLVR—as one unified learning loop. The platform selectively captures only the few high-value interactions that teach something, such as a senior person's correction or a task outcome, and learns from them inside the customer's own environment. This keeps proprietary expertise and team knowledge within the organization rather than handing it to a vendor.
- Artificial Intelligence
- AI Agents
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Most AI systems are trained once before deployment and then remain frozen, unable to improve from real-world usage. They may remember context or past conversations, but they cannot learn from user corrections, feedback, or task outcomes, forcing teams to repeatedly correct the same errors and limiting AI's long-term value.
Solution
Perpetual develops next-generation AI that continuously learns and adapts to the people who use it. The platform uses a unified learning loop that integrates four established techniques—RLHF from user corrections, RLVR from task outcomes, self-rewarding from the model's own judgment, and distillation from organizational knowledge. Rather than storing all data, the system selectively captures only the interactions that matter: the correction a senior person makes that a junior would miss, or the moment where user judgment and model output quietly differ. This learning happens inside the customer's own environment, keeping proprietary expertise and team knowledge within the organization rather than handing it to a vendor.
Target Audience
Engineering and operations teams at organizations deploying AI agents in production who need systems that improve at their specific work and consistently avoid repeating mistakes. Primary customers are enterprises seeking AI that learns from their internal workflows and expertise.
Features
- Unified learning loop treating four signal sources—self-rewarding, RLHF, distillation, and RLVR—as one integrated pipeline
- Selective attention mechanism that filters noisy interaction streams to identify only the few moments that teach something
- Continuous post-deployment learning that improves from how work actually happens, not just from pre-training data
- On-premises or internal learning architecture that keeps organizational knowledge and expertise within the customer's walls
- Agnostic signal intake that learns from model output, user corrections, company knowledge, and task outcomes