Quick Summary
- Trajectory raised $45 million in a Series A led by Sequoia at a reported $300 million valuation.
- Its platform collects feedback from business workflows and uses it to improve models for specialist tasks.
- Customers named by the company include Harvey, Rogo, Decagon, Mercor and Nvidia.
Trajectory has raised $45 million in a Series A round to make the process of improving specialist software models easier for businesses. The startup says its platform can collect feedback from real work and use it to refine models for tasks in fields such as law and finance. TechCrunch reported the funding on November 2, saying the round was led by Sequoia and valued the company at $300 million.
The platform’s intended workflow
- Connect a specialist business model to a real workflow.
- Collect expert feedback on the model’s output.
- Use the feedback to refine performance on a defined task.
- Review the results before deploying an updated model.
The problem is that general models do not automatically learn from each job
Businesses often adapt general-purpose models for specialised work. A legal team may want a system that reviews contracts according to its internal standards, while a financial firm may need software that follows a particular valuation method. Fine-tuning and post-training can improve those results, but the process has traditionally required researchers or specialist consultants who understand how to prepare data and evaluate model behaviour.
Trajectory wants to make that process accessible to people who understand the work but may not have a machine-learning team. The company connects a model to business applications, gathers feedback on how it performs and uses that information to make targeted improvements. The objective is to make a system more useful for a particular organisation over time rather than treating every interaction as a fresh task.
Funding, customers and the competitive field
The 18-person company was founded about a year ago by three Stanford classmates with experience at Google DeepMind and Apple. The Series A was led by Sequoia, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures. TechCrunch reported that the company is making its post-training platform publicly available after working with a group of early customers.
Trajectory named Harvey and Rogo among businesses using its platform to post-train open-source models for legal and financial work. Other customers include Decagon, Mercor and Nvidia. Those relationships provide early evidence that companies are willing to test the approach, but the announcement does not provide audited performance comparisons or a breakdown of revenue by customer. The company competes with providers of fine-tuning platforms and with larger AI labs that provide their own enterprise model tools.
Human review remains central to the pitch
Trajectory describes its approach as continual learning, but the company says people remain involved in deciding what good output looks like. That matters because a model can improve a metric while becoming less useful in ways that are not captured by the training data. Human specialists can define the standards, review the results and decide which changes should be accepted into the next version.
The startup’s founders argue that this approach is safer than allowing a system to modify itself without oversight. It also creates a practical question for customers: how much time will employees need to spend reviewing the feedback loop? If the process requires extensive manual labelling, it may not save as much time as expected. Trajectory’s success will depend on whether its interface makes model improvement repeatable, measurable and less expensive than hiring a specialist team for every new workflow.
Trajectory’s pitch will be tested when customers use the platform on work where errors have a real cost. A legal model that reviews contracts or a financial model that values a company must be evaluated against expert standards, not only user satisfaction. The company will need to show that improvements persist across new examples and do not cause a model to become overly specialised. Its human-feedback approach may help businesses capture internal knowledge, but the quality of the feedback loop depends on who reviews outputs and how disagreements are resolved. The $45 million round gives the startup resources to build that process into a product.
Trajectory’s platform will need to show that model improvements are measurable and repeatable. Customers will want to compare output quality before and after feedback, understand who approved each change and keep their data separate from other organisations. Those controls matter particularly in legal and financial work, where a plausible answer can still be wrong in a consequential way.