Trajectory has raised $45 million to help professionals without technical backgrounds train artificial intelligence models for specialised work, seeking to make their judgement a more direct part of model development. The startup’s platform collects feedback from business applications and uses it to retrain models. Its ambition is to let lawyers, doctors and other subject specialists improve AI systems themselves, rather than leave the process solely to people with technical expertise.

The Series A round was led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures also investing, and valued Trajectory at $300 million. The 18-person company was founded about a year ago by three Stanford classmates whose experience includes Google DeepMind and Apple. Those figures place substantial investor backing behind a young, small business whose proposition is to make the work of adapting models more accessible to the professionals who use them.

Co-founder Ronak Malde describes a distinction between a model’s broad intellectual ability and the experience needed to do a particular job well. He likens foundational models to a highly capable doctoral student arriving in an unfamiliar workplace, while arguing that businesses value the expertise of someone with 30 years in the role. That comparison underpins Trajectory’s approach: the goal is not simply to provide a capable model, but to help it learn from the specific work an organisation needs it to perform.

Read more: European freezer cyberattack reaches UK meat supply chain

The company calls its approach continual learning. A business can begin with an open source model and connect Trajectory’s platform to its applications, enabling the system to gather feedback about performance. Trajectory uses that information for further training, with the aim of improving the model’s handling of specialised assignments over time. The process therefore links model development to experience within the customer’s applications. Feedback from work already performed becomes an input into how the model is subsequently adapted.

Harvey and Rogo are already using the platform to carry out post-training on open source models. The intended result is behaviour closer to that of an experienced lawyer or financial analyst, informed by earlier assignments. Examples include examining a particular kind of legal contract and applying a specific methodology to value a company. These uses illustrate the narrow, organisation-relevant expertise Trajectory is targeting, rather than a general improvement claim across every kind of task an AI system might encounter.

Other customers include Decagon, Mercor and Nvidia, giving the chip company a place among both Trajectory’s users and its investors. The customer list provides examples of adoption alongside the financing figures, although those are different measures from the performance improvements the platform seeks to deliver. The $45 million round and $300 million valuation indicate investor support for the business. The operational proposition rests on whether feedback can help customers make models better suited to the particular tasks they want automated.

Read more: U.S. strikes Iranian targets in Strait of Hormuz as Tehran asserts self-defense right

Malde says specialists can use Trajectory to improve models with much of the process handled automatically, potentially reducing the time and cost of deploying specialised AI. Users determine the standard of performance their organisation needs and feed that judgement into model improvement. This puts professional expertise at the centre of the product’s appeal: the people who understand a contract review or a valuation methodology can help shape the model without needing to manage all the technical steps of post-training themselves.

That emphasis on human judgement also marks a distinction from recursive self-improvement, an idea under which AI systems develop their own capabilities. Trajectory’s continual-learning approach has similarities to that concept, but co-founder Michael Elabd argues that retaining human involvement makes its method safer. In his view, autonomous improvement could produce unintended or dangerous results. The contrast is between allowing systems to direct their own development and using people’s assessments to guide it, rather than a rejection of automation within training altogether.

Malde regards the team as the company’s principal advantage. He was among researchers brought into Google DeepMind through its acqui-hire of Windsurf personnel, while Elabd also worked at DeepMind. The third co-founder, Arjun Karanam, previously worked on generative foundation models at Apple. Trajectory has additionally hired researchers and product designers from OpenAI, Stripe and Netflix. That combination reflects the company’s dual focus on the technical work of model training and the interface through which customers provide the expertise needed to guide it.

Karanam says the company has built an interface intended to make post-training easier for nontechnical users. That design effort connects the financing announcement to Trajectory’s broader objective: bringing the capacity to adapt models closer to the specialists who can assess their output. Its stated direction is continued improvement through feedback on real assignments. Whether that approach delivers faster, less costly deployment remains the practical test of its proposition, with human participation positioned as both a product feature and a safeguard.