A group of artificial intelligence researchers who previously worked at Google DeepMind, Apple, OpenAI and Meta Superintelligence Labs announced Wednesday that they are launching a fresh startup called Trajectorywhich aims to aid companies regularly improve their AI products through real-world user interaction training.
Trajectory wants to build a platform for artificial intelligence that can learn continuously, which researchers have long considered a major obstacle to further progress in the field of artificial intelligence. OpenAI, Google and Anthropic have demonstrated success in training increasingly productive versions of AI models, especially in domains such as coding, math and science. However, these systems stop getting smarter after training. While there have been some recent breakthroughs in continuous learning, technology companies have generally struggled to create AI products that learn from their mistakes in real time. In December 2025 at NeurIPS, one of the largest annual research conferences on artificial intelligence, Turing Award winner Richard Sutton argued that continuous learning is essential to build superintelligent agents.
Trajectory raised a $15 million seed round at a post-money valuation of $115 million, led by venture capital firm Conviction, with participation from Bessemer Venture Partners, Radical VC and BoxGroup. Individual investors also participated in the round, including: Google DeepMind chief scientist Jeff Dean, as well as the so-called the “godmother of AI,” Stanford professor and World Labs CEO Fei-Fei Li.
Trajectory CEO and co-founder Ronak Malde was previously an artificial intelligence researcher at Windsurf and later became one of the few employees to join Google DeepMind when it hired the most talented person from the coding startup at A deal worth $2.4 billion last year. Trajectory’s other co-founders are Arjun Karanam, a former artificial intelligence researcher at Apple who worked on Vision Pro, and Michael Elabd, who previously worked at Google’s DeepMind robotics division.
Malde tells WIRED that some leading AI coding products, like Cursor, are already using an early version of continuous learning — using real data about how people interact with your products to conduct post-training training and send model improvements regularly. He argues that this is the main reason why AI coding products have grown so rapidly in popularity, and one of the reasons why major AI labs have rushed to develop their own vibration coding applications. With Trajectory, Malde and his team of 11 researchers and engineers hope to apply a similar technique to improve AI-based tools beyond the coding space.
“Even the most powerful AI today is still static. The AI model you used yesterday will make the same mistakes today,” says Malde. “Several companies are starting to enter the world of continuous learning. We are building a platform for every company to learn continuously.”
The challenge with applying this logic to other fields is that coding is easily verifiable – the code either works or it doesn’t – but in some industries there is a looser definition of success. Karanam says the Trajectory platform, in part, helps optimize the AI model for a company’s specific needs.
Instead of starting with an off-the-shelf OpenAI or Anthropic model, Trajectory offers customers to start with an open source model that has been trained for the specific AI product the company has in mind. For Decagon, an AI-powered customer service agent developer, Trajectory records when the AI doesn’t live up to its expectations — say, an inquiry from a customer trying to make a return is sent back to a human — and uses those instances to later train a fresh model on an up to weekly basis. Trajectory claims that these trained models outperform frontier lab models on the narrow tasks that matter most to a company’s product.