Trajectory raises $45 million to put model fine-tuning in experts’ hands

Turning feedback into model updates

Trajectory’s approach is a form of continuous learning: a company connects an open model to its applications, the platform collects feedback on its performance, and that feedback informs further training.

Co-founder Ronak Malde says users define “what a good result looks like for their organization,” then turn that standard into model improvements. The goal is to help models handle the specific work of a business, from reviewing a type of legal contract to evaluating a company using a particular methodology.

Harvey and Rogo already use the platform to further train open models on past tasks, making them behave more like experienced lawyers or financial analysts. Other customers include Decagon, Mercor and Nvidia. Trajectory is now opening the platform to the public.

The round also included Nvidia, Bessemer Venture Partners and Bain Capital Ventures.

Human feedback as a safety boundary

Continuous learning has some resemblance to recursive self-improvement: the idea that AI systems could develop their own capabilities. Co-founder Michael Elabad argues that keeping people involved in training makes Trajectory’s approach safer, because autonomous self-improvement could produce unforeseen, even dangerous, outcomes.

Trajectory is not alone in trying to automate further model training. The company sees Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, as a direct competitor; it built the fine-tuning platform Tinker. Other competitors include Applied Compute, valued at $3.25 billion last month, and Prime Intellect, which helps companies build and train their own AI agents.

The interface is part of the product

Malde points to the team as Trajectory’s main advantage. He was among several Google DeepMind researchers hired alongside Windsurf assets. Elabad also worked at DeepMind, while third co-founder Arjun Karanam worked on generative foundation models at Apple. The startup has also hired researchers and product designers from OpenAI, Stripe and Netflix.

Karanam’s case for product design is straightforward: customers without technical training need an interface they can use to work with model fine-tuning.

I think that is the more consequential part of the pitch. Trajectory is not just trying to make training easier; it is betting that specialists can reliably decide what “good” means for their work and feed that judgment back into a model. The announcement says little about how those judgments are checked or how conflicting feedback is handled. If the interface makes fine-tuning accessible but leaves those decisions opaque, the hard part has not disappeared—it has moved closer to the user.