Trajectory, a startup founded by three former Stanford classmates with experience at Google DeepMind and Apple, has secured $45 million in Series A funding, bringing its valuation to $300 million. The company’s mission is to simplify AI model training for businesses by making it accessible to nontechnical users. Trajectory’s technology focuses on continual learning, allowing AI models to improve over time by learning from ongoing feedback, which is particularly useful for specialized fields like law and finance where fine-tuning AI models usually requires costly experts.

The platform works by enabling clients to start with an open source AI model and connect it to their applications to collect performance feedback. Trajectory then retrains the AI model based on this data, gradually enhancing its proficiency in specific tasks. This approach allows professionals such as lawyers and doctors to oversee AI improvements directly, reducing the need for deep technical expertise and speeding up the adoption of customized AI solutions. Early customers of Trajectory include companies like Harvey, Rogo, Decagon, Mercor, and Nvidia.

Trajectory’s co-founders emphasize the safety advantage of their system by maintaining human involvement in the AI training process, distinguishing it from recursive self-improving AI systems that operate autonomously and could pose risks. The startup is part of a growing sector targeting the automation of AI fine-tuning, with competitors including Thinking Machines Lab, Applied Compute, and Prime Intellect. The company credits its rapid progress to the strong technical pedigree of its team, which includes researchers with backgrounds at OpenAI, Stripe, Netflix, and previous work on foundation models at Apple.

By designing a user-friendly interface, Trajectory aims to demystify post-training AI refinement and empower businesses without large AI research teams to benefit from advanced machine learning. The startup is now making its post-training platform broadly available, allowing organizations to define their own success criteria and translate them into ongoing AI improvements. This strategy could lower barriers to entry for AI customization, enabling wider deployment of specialized AI models that adapt and improve continuously with minimal technical intervention.