Ex-Google and Apple Researchers Start New Venture to Create AI’s Essential Feedback Loop

A collective of AI researchers with prior experience at Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs announced on Wednesday the launch of a new startup named Trajectory, designed to assist companies in consistently enhancing their AI products by leveraging real-world user interactions.

Trajectory’s goal is to create a platform for AI that can learn continuously, addressing a significant challenge that has hindered advancements in the field. While OpenAI, Google, and Anthropic have successfully developed increasingly proficient AI models—especially in areas like coding, math, and science—these systems cease to evolve once their training concludes. Despite some recent breakthroughs in continual learning, tech firms have generally struggled to develop AI products that can learn from mistakes in real time. During the NeurIPS conference in December 2025, esteemed researcher Richard Sutton emphasized that continual learning is crucial for crafting superintelligent agents.

Trajectory has secured a $15 million seed round at a post-money valuation of $115 million, led by the venture capital firm Conviction, with contributions from Bessemer Venture Partners, Radical VC, and BoxGroup. Noteworthy individual investors include Google DeepMind’s chief scientist, Jeff Dean, and Stanford professor and World Labs CEO Fei-Fei Li, who is often referred to as the “godmother of AI.”

Ronak Malde, Trajectory’s CEO and cofounder, previously conducted AI research at Windsurf and was among a select few employees recruited by Google DeepMind following a $2.4 billion acquisition of the coding startup’s top talent last year. His fellow cofounders are Arjun Karanam, an ex-AI researcher at Apple who contributed to the Vision Pro, and Michael Elabd, who worked in Google DeepMind’s robotics division.

Malde tells WIRED that leading AI coding products, like Cursor, are already implementing a preliminary form of continual learning—using real data to refine their models and deliver regular improvements post-training. He suggests this is a core factor behind the rapid growth of AI coding tools and a driving reason for major AI labs rushing to create their own coding applications. With Trajectory, Malde and his team of 11 researchers and engineers intend to utilize a similar method to enhance AI-powered tools beyond the coding sector.

“Even the most advanced AI today remains static. The AI model you utilized yesterday will repeat the same errors today,” Malde states. “A few companies are moving towards continual learning. Our aim is to establish a platform enabling every company to achieve that.”

Applying this concept to various domains is challenging since coding is easily verifiable—code either executes or it does not—while other industries may have more ambiguous definitions of success. Karanam explains that a key aspect of Trajectory’s platform is its ability to optimize an AI model according to a business’s specific requirements.

Instead of launching from a generic model from OpenAI or Anthropic, Trajectory starts its clients with an open-source model prepped for a particular AI application in mind. For example, Decagon, a client developing AI customer support agents, relies on Trajectory to track shortcomings—such as when a customer’s return inquiry is redirected to a human—and uses these instances to continually post-train a new model, achieving updates potentially every week. Trajectory asserts that these post-trained models outperform those from leading labs on the specific tasks that are most critical to a company’s product.