Trajectory raises $45M to automate AI training for nontechnical experts

Bridging the expertise gap

Current artificial intelligence models often struggle with specialized tasks because they do not naturally learn from their mistakes after their initial training. While companies in fields like finance or law can use fine-tuning to adapt these models, the process is usually difficult and costly. It typically requires hiring internal research teams or expensive external consultants known as forward-deployed engineers.

Trajectory is a startup founded by three former employees of Google DeepMind and Apple who want to automate this complex workflow. Their goal is to create a system where AI models can learn from feedback and improve continuously at specific jobs. By making this process accessible, they hope to allow non-technical experts to manage their own AI tools.

The company recently announced it raised $45 million in a Series A funding round. This investment was led by Sequoia and included participation from major tech players like Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.

How continual learning works

The core technology behind Trajectory is an approach called continual learning. This allows a model to receive feedback on its performance and then be retrained to incorporate that information. Instead of a one-time training session, the model can evolve as it interacts with real-world data and user input.

Businesses can start by using an open-source AI model and connecting it to their existing applications. As the model performs tasks, the platform collects data on how well it succeeded or failed. Trajectory then uses that specific data to retrain the model, helping it get better at specialized tasks over time.

Co-founder Ronak Malde described foundational models as being like a very smart PhD student who is always starting a new job. He noted that the economy values someone who has worked on the job for 30 years, highlighting the need for models that can gain deep, practical experience.

Impact on specialized industries

Trajectory is already being used by companies like Harvey and Rogo to post-train open-source models. These firms use the platform to help their AI perform like experienced lawyers or financial analysts. For example, the models can learn to review specific types of legal contracts or value companies using very particular methodologies.

Other customers using the platform include Decagon, Mercor, and Nvidia. The startup's goal is to make it possible for doctors, lawyers, and other specialists to improve their models nearly automatically. This could significantly decrease the time and money required to deploy high-quality, specialized AI tools.

The platform is now being made publicly available to a wider audience. This move aims to let users decide what good performance looks like for their specific organization. Users can then turn those specific standards into direct improvements for their custom models.

Safety and human oversight

Trajectory's approach to continual learning shares some similarities with recursive self-improvement. This is the concept where AI systems might be able to improve their own capabilities without human intervention. However, the founders argue that keeping humans in the loop is a critical safety feature.

Co-founder Michael Elabd believes that human involvement makes the training process safer. He suggests that allowing an AI to improve itself autonomously could lead to unintended or even dangerous outcomes. By keeping the human expert in charge of the feedback, the system remains aligned with human goals.

The startup is also navigating a competitive landscape in the post-training space. They view Thinking Machines Lab, Applied Compute, and Prime Intellect as direct competitors. Trajectory aims to differentiate itself through its team's deep experience and its focus on accessible design.

Design for non-technical users

A key part of Trajectory's strategy is making complex AI training accessible to people who are not engineers. Co-founder Arjun Karanam noted that the startup developed a specific interface to achieve this. This design allows non-technical customers to interact with the training process without needing to write complex code.

The team behind the company includes researchers and product designers from notable companies like OpenAI, Stripe, and Netflix. This diverse background helps them balance the technical requirements of AI training with the practical needs of the end user. They want to ensure that the tool is usable for the people who actually understand the subject matter, such as legal or medical professionals.

By focusing on the user experience, Trajectory hopes to democratize the ability to create specialized AI. This shift could move the power of AI development away from just the largest tech companies and into the hands of domain experts. This transition could define how specialized AI is built and deployed in the coming years.