AI models can handle many tasks, but they do not automatically learn from mistakes made while doing specialized work. Trajectory wants to make it easier for businesses to turn human feedback into improvements, without relying on large internal research teams or expensive consultants.
Turning work feedback into model training
Businesses adapting general AI models for fields such as law and finance often fine-tune those models for their needs. That work can require dedicated researchers or forward-deployed engineers, adding cost and making the process harder to manage.
Trajectory’s platform takes a different route: a business starts with an open-source model and connects it to its applications. The platform collects feedback on how the model performs, then uses that information to retrain it for specialized tasks.
The approach is called continual learning. Rather than treating training as a one-time step, it gives organizations a way to incorporate feedback from ongoing use. The stated aim is for models to get better at tasks that matter to a particular business.
People define what a good result means
Trajectory co-founder Ronak Malde described foundational models as a very smart PhD student who is always starting a new job. He told TechCrunch, “But what the economy values is someone who worked on the job for 30 years.”
That comparison points to the gap the company is trying to address: broad capability does not necessarily mean familiarity with the specific work, standards, or methods an organization needs. Trajectory says specialists can guide improvement themselves by deciding “what good looks like for their organization and can turn that into improvements for their model,” as Malde put it.
In practice, the feedback loop could help a model learn from repeated tasks. The article gives examples including reviewing a particular type of legal contract and valuing a company with a specific methodology. Those examples show how a general model might be adapted to more consistent, domain-specific work.
Customers and a newly public platform
Companies including Harvey and Rogo use Trajectory to post-train open-source models for work associated with experienced lawyers or financial analysts. Other customers named in the article are Decagon, Mercor, and Nvidia. Trajectory is now making its post-training platform publicly available.
The company argues that allowing lawyers, doctors, and other specialists to improve models more directly could make specialized AI faster and less expensive to deploy. The platform’s interface is part of that effort: co-founder Arjun Karanam says it is designed to make post-training more accessible to non-technical customers.
That accessibility matters because a feedback system is only useful to a business if the people who understand the work can take part. Trajectory’s model puts those specialists in the loop, while the platform handles retraining based on the information they provide.
Funding, competition, and human oversight
Trajectory, founded about a year ago by three Stanford classmates, announced a $45 million Series A at a $300 million valuation. Sequoia led the round, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. The company has 18 employees.
It is entering a field with other efforts to automate post-training. Trajectory considers Thinking Machines Lab, Applied Compute, and Prime Intellect competitors. The article describes Thinking Machines Lab as the company behind the fine-tuning platform Tinker, and says Applied Compute was valued at $3.25 billion last month.
Trajectory’s founders also distinguish their approach from recursive self-improvement, the idea that AI systems can improve their own capabilities. Co-founder Michael Elabd says keeping people involved in training makes the approach safer, because autonomous self-improvement could produce unintended or dangerous outcomes.
The company’s bet is that specialized AI can improve when feedback from real work is turned into training, and that the process can be made usable by people without technical training. Its funding and customer list provide a starting point; whether the approach can make this kind of model improvement routine across organizations will depend on how well the feedback loop works in practice.