Whether developed by OpenAI, Anthropic, or built on open-source alternatives, current AI models have a key limitation: without further training, they cannot learn from their mistakes.
Companies in professional fields such as legal and financial services typically address this issue by fine-tuning general-purpose models. However, this often requires hiring an internal research team or relying on expensive consultants, commonly known as front-end deployment engineers (FDEs).
Trajectory, a startup founded by three Stanford classmates about a year ago, aims to automate this process, enabling AI models to learn from feedback and continuously improve on specialized tasks. All three founders previously worked at Google DeepMind and Apple.
The company, which has 18 employees, has just announced the completion of a $450 million Series A round at a $3 billion valuation. The round was led by Sequoia, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.
Ronak Malde, co-founder of Trajectory, compared the foundational model to "a very smart PhD student who always starts a new job." "But what the economy truly values is someone who has worked in the role for 30 years," he told TechCrunch.
This startup aims to address this issue through a method called continuous learning. Companies using Trajectory can start with an open-source AI model and integrate the platform into their own applications to collect feedback on the model’s performance. Trajectory then uses this data to retrain the model, helping it continuously improve on specialized tasks over time.
Malde said that with Trajectory, lawyers, doctors, and other professionals can almost automatically improve their own models. This capability could make deploying specialized AI faster and cheaper.
Companies like Harvey and Rogo are already using Trajectory’s platform to fine-tune open-source models, helping them work more like experienced lawyers or financial analysts by enabling the models to learn from past tasks—such as reviewing certain types of legal contracts or valuing companies using specific methodologies. Other Trajectory customers include Decagon, Mercor, and Nvidia. The startup is now opening its fine-tuning platform to the public.
Malde said that users of Trajectory will decide "what constitutes good for their organization and can translate that into improvements to the model."
Trajectory’s continuous learning approach shares some similarities with recursive self-improvement (RSI), which refers to AI systems enhancing their own capabilities. However, co-founder Michael Elabd believes that involving humans in the training process makes Trajectory’s approach safer, as autonomous self-improvement could lead to unintended or even dangerous outcomes.
Trajectory is not the only startup attempting to automate post-training. The company considers Thinking Machines Lab, founded by former OpenAI Chief Technology Officer Mira Murati and known for launching the fine-tuning platform Tinker, as a direct competitor, along with Applied Compute, which reached a $3.25 billion valuation last month, and Prime Intellect, which helps companies build and train their own AI agents.
Malde believes the startup’s greatest strength lies in its team. He was one of the few researchers recruited from Windsurf by Google DeepMind. Elabd also previously worked at DeepMind, and Trajectory’s third co-founder, Arjun Karanam, previously led generative foundation models at Apple. The company has also recruited researchers and product designers from OpenAI, Stripe, and Netflix.
Why does a company focused on AI training place such strong emphasis on design? Karanam explains that the company developed an interface that makes post-training more accessible for its non-technical customers.