Startup Trajectory Secures $45M to Democratize AI Model Fine-Tuning

Current AI models, regardless of whether they are from OpenAI, Anthropic, or open-source alternatives, suffer from a significant flaw: they struggle to learn from their errors without undergoing additional training.

To adapt these general models for specialized sectors like law and finance, businesses typically rely on fine-tuning. However, this process usually necessitates hiring internal research teams or engaging expensive consultants, commonly known as forward-deployed engineers (FDEs).

Trajectory, a startup established about a year ago by three Stanford alumni—each with backgrounds at Google DeepMind and Apple—seeks to automate this process. The company enables AI models to learn from feedback and continuously enhance their performance on specific tasks.

The 18-person company recently announced a Series A funding round of $45 million at a valuation of $300 million. The round was led by Sequoia, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.

Ronak Malde, a co-founder of Trajectory, drew a parallel between foundational models and a brilliant PhD student starting a new job. "However, what the economy truly values is someone who has dedicated thirty years to that specific role," he explained to TechCrunch.

The startup intends to address this challenge through a technique known as continual learning. By connecting their platform to existing applications, businesses can utilize Trajectory to gather performance feedback. The platform then utilizes this data to retrain the model, facilitating long-term improvement on specialized tasks.

Malde noted that with Trajectory, professionals such as lawyers and doctors can enhance models with minimal manual effort. This capability has the potential to significantly accelerate and reduce the cost of deploying specialized AI solutions.

Organizations like Harvey and Rogo are already utilizing Trajectory to post-train open-source models, enabling them to mimic the expertise of experienced lawyers or financial analysts by learning from specific tasks like contract review or valuation methodologies. Other customers include Decagon, Mercor, and Nvidia, with the startup now offering its post-training platform to the public.

"The people using Trajectory are the ones who decide what success looks like for their organization and can translate that into model improvements," Malde stated.

Trajectory's approach to continual learning shares some parallels with recursive self-improvement (RSI), the concept that AI systems can enhance their own capabilities. Nevertheless, co-founder Michael Elabd argues that maintaining human involvement in the training loop makes Trajectory's method safer, as autonomous self-improvement could result in unintended or dangerous consequences.

Trajectory is not the only startup attempting to automate post-training. The company considers Thinking Machines Lab, founded by former OpenAI CTO Mira Murati and the creator of the Tinker fine-tuning platform, a direct competitor. Others in the space include Applied Compute, valued at $3.25 billion last month, and Prime Intellect, which assists companies in constructing and training their own AI agents.

Malde believes the startup's primary advantage lies in its team. He was part of the group of researchers Google DeepMind acquired from Windsurf. Elabd also has a background at DeepMind, while the third co-founder, Arjun Karanam, previously focused on generative foundation models at Apple. The team has also recruited talent from OpenAI, Stripe, and Netflix. Karanam explained that design is crucial because the startup has developed an interface that simplifies post-training for non-technical clients.