Former Google, Apple Researchers Launch AI Feedback Loop Startup

A team of former researchers from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs has officially unveiled a new startup, Trajectory, designed to help companies continuously improve their artificial intelligence products by learning from real-world user interactions. The announcement was made Wednesday.

The core mission of Trajectory is to build a platform enabling continuous learning for AI systems, a long-standing hurdle in the field. While companies like OpenAI, Google, and Anthropic have made impressive strides in training increasingly powerful models, especially for coding, math, and science, these systems tend to plateau after their initial training is complete. Despite recent advances, tech firms have largely failed to create AI products that learn from their mistakes in real time. At the NeurIPS conference in December 2025, Turing Award winner Richard Sutton argued that achieving continual learning is critical for developing superintelligent agents.

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

The startup’s CEO and cofounder, Ronak Malde, previously worked as an AI researcher at Windsurf, a coding startup later acquired by Google DeepMind in a $2.4 billion deal last year. Malde was among the few employees who transitioned to DeepMind as part of that acquisition. His cofounders include Arjun Karanam, a former Apple AI researcher who contributed to the Vision Pro, and Michael Elabd, who worked in Google DeepMind’s robotics division.

Malde told WIRED that some leading AI coding tools, such as Cursor, already employ an early form of continual learning. They use real user interaction data to perform post-training and regularly ship model improvements. He believes this is a key factor behind the rapid adoption of AI coding products and has driven major AI labs to develop their own vibe coding applications. With Trajectory, Malde and his team of 11 researchers and engineers aim to extend this technique to AI-powered tools beyond coding.

“Even the most powerful AI today is still static. The AI model that you used yesterday is going to make the same mistakes today,” Malde explained. “A couple companies are starting to get to that world of continual learning. What we are doing is building the platform for every single company to get to continual learning.”

Applying this approach to other domains poses a challenge because coding is easily verifiable: code either runs or it does not. Other industries have looser definitions of success. Karanam noted that part of Trajectory’s value lies in helping optimize an AI model to a business’s specific needs.

Instead of starting with a pre-built model from OpenAI or Anthropic, Trajectory guides customers to begin with an open-source model that has been post-trained for a particular AI product. For example, Decagon, a customer that builds AI customer support agents, uses Trajectory to log instances where its AI falls short, such as when a customer trying to initiate a return gets transferred to a human agent. These instances are then used to post-train a new model, sometimes as often as every week. Trajectory claims these post-trained models outperform frontier labs’ models on the narrow tasks most critical to a company’s product.