Trajectory Raises $45M Series A at $300M Valuation to Make AI Model Training Accessible

AI startup Trajectory has raised $45 million in a Series A funding round at a $300 million valuation, as it looks to make specialised AI model training more accessible to businesses and nontechnical experts.

The round was led by Sequoia, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures. Founded approximately a year ago by three Stanford classmates, the company is developing a platform that allows businesses to continuously improve AI models using feedback from real-world tasks.

Based on the approach known as continual learning, Trajectory aims to reduce the need for businesses to hire dedicated AI research teams or expensive consultants to customise models for specialised applications.

Making AI models learn from feedback

Today's general-purpose AI models can perform a wide range of tasks, but they do not automatically learn from their mistakes or improve their performance through everyday use. Businesses operating in specialised fields such as law and finance often need models that understand their specific workflows, standards and expectations.

One common solution is fine-tuning, which involves adapting an existing model using additional training data. However, this process can require specialised technical teams or external forward-deployed engineers.

Trajectory is building a platform to automate much of this process.

Businesses can connect an open-source AI model to their applications through the platform. Trajectory collects feedback on the model's performance and uses that information to retrain it, helping the model improve at specific tasks over time.

The objective is to allow domain experts, including lawyers and doctors, to contribute to model improvements without needing to become AI researchers themselves.

Co-founder Ronak Malde compares foundational AI models to highly intelligent graduates who repeatedly start new jobs. While they may have strong general knowledge, businesses need systems that develop the experience and expertise required to perform specialised work consistently.

Used by companies across AI and enterprise software

Trajectory's platform is already being used by companies working in specialised AI applications.

Customers include Harvey, a legal AI startup, and Rogo, which develops AI tools for financial professionals. These companies use the platform to post-train open-source models so they can perform specialised tasks more effectively.

Examples include reviewing particular types of legal contracts and valuing companies according to specific financial methodologies.

Other customers include Decagon, Mercor and Nvidia.

The startup is now making its post-training platform publicly available, potentially opening access to a broader range of businesses looking to customise AI models for their own workflows.

Rather than relying entirely on general-purpose model capabilities, organisations can define what good performance means for their operations and use feedback to improve the models accordingly.

A human-guided approach to continual learning

Trajectory's approach shares similarities with recursive self-improvement, a concept in which AI systems improve their own capabilities.

However, co-founder Michael Elabd says the company's approach keeps humans involved in the training process. Specialists determine what constitutes a good outcome, and their feedback helps guide model improvements.

This human involvement is intended to reduce the risks associated with fully autonomous self-improvement, where systems could make unintended changes to their own capabilities.

The company is positioning its platform as a way to combine continuous model improvement with human oversight.

Founders bring experience from Google DeepMind and Apple

Trajectory was founded by Ronak Malde, Michael Elabd and Arjun Karanam, who met while studying at Stanford.

Malde was among the researchers acquired by Google DeepMind through its acqui-hire of Windsurf. Elabd also previously worked at DeepMind, while Karanam worked on generative foundation models at Apple.

The startup has also recruited researchers and product designers from companies including OpenAI, Stripe and Netflix.

The founding team believes product design is central to making advanced AI training accessible to people without technical backgrounds. Its interface is designed to allow domain experts to participate in post-training without navigating complex machine-learning workflows.

Trajectory currently has 18 employees.

Competition in the AI post-training market

Trajectory is entering a competitive market as more companies develop tools to help businesses customise, fine-tune and train AI models.

The company identifies Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, and its fine-tuning platform Tinker as a direct competitor. Other competitors include Applied Compute and Prime Intellect, which helps businesses build and train their own AI agents.

As businesses increasingly deploy AI in specialised workflows, the ability to improve model performance using real-world feedback could become an important part of enterprise AI infrastructure.

With $45 million in fresh funding, a $300 million valuation and customers across legal, finance and enterprise AI, Trajectory is looking to make continuous model improvement more accessible beyond specialist research teams.