Trajectory Completes $45 Million Series A Funding to Automate AI Model Fine-Tuning Process

·by Henderson·AI
Trajectory Completes $45 Million Series A Funding to Automate AI Model Fine-Tuning Process

Regardless of whether the model comes from OpenAI, Anthropic, or is adapted from open-source solutions, current AI technology shares a common shortcoming: without additional training, it cannot learn from its own mistakes. If companies want to apply general models in specialized fields such as law and finance, fine-tuning is usually necessary, which often means building internal research teams or hiring expensive external deployment engineers.

Founded about a year ago, the startup Trajectory aims to automate this process. The company's three founders are Stanford alumni who previously worked at Google DeepMind and Apple. The company recently announced the completion of a $45 million (approximately 320 million RMB) Series A funding round, achieving a valuation of $300 million (approximately 2.13 billion RMB), led by Sequoia Capital, with participation from NVIDIA, Bessemer, and Bain Capital Ventures. The company currently has only 18 employees.

Trajectory Offers Automated AI Fine-Tuning Solutions

Trajectory adopts a continuous learning approach: companies can start with an open-source model, integrate the platform into their applications, and allow the system to continuously collect feedback on model performance, using this information to retrain the model to improve its performance on specific tasks. Co-founder Ronak Malde likens the foundational model to a very intelligent but perpetually job-changing PhD student, "while the economy truly values someone who has worked in one position for 30 years."

According to Malde, with Trajectory, professionals such as lawyers and doctors can essentially fine-tune the model themselves without relying on algorithm teams, making the deployment of specialized AI faster and cheaper. Harvey and Rogo are already using this platform to retrain open-source models to behave more like experienced lawyers or financial analysts, such as learning to review a specific type of contract or valuing a company using particular methods. Other clients include Decagon, Mercor, and NVIDIA, and the platform is now publicly available.

This approach is similar to "recursive self-improvement," but co-founder Michael Elabed believes that involving humans in the training process makes Trajectory safer, as fully autonomous self-improvement could lead to unexpected or even dangerous outcomes. There are numerous competitors in this field, including the fine-tuning platform Tinker from Thinking Machines Lab, Applied Compute, which was valued at $3.25 billion (approximately 23.1 billion RMB) last month, and Prime Intellect, which helps companies manage their self-built intelligent agents, all viewed as direct competitors.

Malde views the team as the company's greatest asset: he himself was one of the researchers recruited from Windsurf to Google DeepMind, Elabed also comes from DeepMind, and the third founder, Arjun Kalanan, previously worked on generative foundational models at Apple. The company has also recruited research and product design personnel from OpenAI, Stripe, and Netflix. Kalanan stated that they have specifically designed an interface that allows non-technical clients to easily engage in fine-tuning.

H
About the author
Henderson