DynamoFL

Dynamo operates a platform that allows developers to train AI models on personally identifiable information (PII) without compromising privacy, at scale. The company aims to make AI development more accessible in industries where data privacy is critical, such as healthcare and financial services. Dynamo is a graduate of Y Combinator’s 2022 winter cohort and was originally founded in 2021 by two PhD graduates from the Massachusetts Institute of Technology (MIT). 

The Dynamo solution is composed of three modules: i) DynamoEval for automated stress testing of AI systems; ii) DynamoEnhance, which remediates identified risks and enhances models to improve security and privacy; and iii) DynamoGuard to enable enterprises to deploy customizable AI guardrails.

At its core, Dynamo AI’s platform uses federated learning, a method that aggregates several smaller AI models trained by hundreds or thousands of different users in their own environments to create the final full-fledged model, removing the need to share sensitive data. Federated learning is, however, considered difficult to deploy, with several drawbacks, including costs associated with transferring AI models back and forth, inaccuracy of the final model due to statistical variation across the component models’ data, and lack of personalization.

Dynamo has developed a proprietary system, “FedLTN,” which avoids these drawbacks by using various techniques to prune AI models to a smaller size while maintaining accuracy. The platform also allows users to fine-tune models to specific cohorts.