Trajectory raises $45M: continual learning for open models

Every AI model you rent today has the same flaw: it does not learn from the mistakes it makes in your business. You correct it on Monday, and it makes the same mistake on Tuesday. Trajectory, an 18-person startup, just raised a $45 million Series A at a $300 million valuation to sell continual learning as a product: connect an open-source model to your app, collect feedback, and retrain it on that feedback. TechCrunch reported the round on November 2.

What actually happened

From TechCrunch's report:

• The round: $45M Series A led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures. Valuation: $300M. The company is about a year old.
• The product: a business starts with an open-source model and connects Trajectory's platform to its application. The platform collects feedback on how the model performs, then uses that feedback to post-train the model on the specialized task.
• Who it is for: the pitch is that lawyers, doctors and other specialists decide "what good looks like" and drive the improvements themselves, through an interface built for non-technical users. No in-house research team, no forward-deployed engineers.
• Customers: Harvey and Rogo use it to post-train open models for legal and financial work. Decagon, Mercor and Nvidia are also listed as customers.
• Availability: the post-training platform is now publicly available.
• Competition: Trajectory names Thinking Machines Lab (maker of the Tinker fine-tuning platform), Applied Compute and Prime Intellect as competitors.

One note: earlier reports put this round at $40M. We use TechCrunch's $45M figure, which TechCrunch attributes to the company's own announcement.

Why it matters for your business

Continual learning only works with one input: a record of where the model was wrong and what right looked like. Most small businesses running AI today throw that record away. A staff member fixes an AI-drafted quote, edits a product description, or rewrites a support reply, and the fix goes out to the customer. The model never sees it.

That record is the asset, whatever vendor you pick later. If you log every human correction now (the input, the model's output, the edited version, who approved it), you own a training set built from your actual work. You can use it to post-train an open model on Trajectory, on a competitor, or in-house. You can also use it today, without any training, as an eval set to test whether a new model handles your edge cases before you switch.

The open-source angle matters too. This works because the customer owns the weights it trains. You cannot post-train a closed API model the same way, so the improvement stays with the vendor. If your AI workflow depends on one closed model, the knowledge your team puts into it does not stay with you.

Sources: TechCrunch.