Trajectory Raises $45M Series A for Continual AI Learning

A rejected AI answer can contain something worth keeping: the reason an experienced person rejected it. Trajectory is building a business around carrying that judgment into a model's next training cycle, rather than leaving it inside an individual correction.

The San Francisco company announced a $45M Series A on November 2, 2026, led by Sequoia at a $300M valuation, with NVIDIA, Bessemer and Bain Capital Ventures participating. Co-founders Ronak Malde, Michael Elabd and Arjun Karanam are taking post-training into a product that combines training, evaluation and deployment. The financing arrives as the company opens its platform and introduces research previews aimed at making expert feedback more reusable.

For operators, the useful question is how knowledge becomes part of the software an organization runs. Trajectory's approach gives subject experts a place in that process, while leaving teams responsible for defining success and deciding when an update should reach production.

What Trajectory's Series A finances

Trajectory's new round is $45M; the company labels its combined Seed and Series A funding as $60M. The $300M disclosed valuation has no stated pre-money or post-money basis in the official announcement. The company has not published a detailed allocation of the proceeds, so a precise budget for hiring, research or expansion would go beyond the disclosure.

The commercial ambition is visible in the product launch. A customer brings tasks and a way to assess them, then works through training, evaluation and deployment inside one environment. Packaging those activities together matters because each handoff can otherwise require a team to reconcile different tools, definitions and records. That is an operating interpretation of the product, rather than a disclosed saving every customer can expect.

Bessemer's portfolio description places Trajectory in enterprise post-training based on product-usage telemetry. The investment is therefore close to a continuing customer activity: collecting evidence about how an AI application performs after people begin using it. The value of that evidence depends on whether it can change the next version of the product in a useful, measurable way.

The expert's standard becomes training material

Continual Grading, a research preview, addresses a problem that appears when a task has no simple automatic answer key. Legal analysis and financial work can require judgment about the quality of an answer, with standards that differ between organizations. Asking experts to write criteria for every training example creates a substantial labeling assignment alongside their existing jobs.

Trajectory's reported method uses a small collection of expert-graded tasks to learn a procedure that generates grading criteria for new tasks. The company calls that procedure a metagrader. Its purpose is to carry something learned from the expert examples into unfamiliar work, reducing the amount of task-by-task rubric writing the customer must supply.

That makes the selection of examples a consequential management decision. A team needs to know which judgments represent its standards and how disagreements should be handled before those judgments influence training. This is analysis of the method's organizational implications, rather than a claim that the software resolves every disagreement. An interface can make contribution easier while leaving the underlying choice firmly with the business.

A learning loop with deployment decisions

The platform documentation describes a spec as the definition of what an agent should improve at, with tasks and grading criteria used to measure a candidate model. The workflow includes evaluating a base model, training, comparing a checkpoint on held-out tasks and choosing a model to serve. Those separate steps give an operator places to inspect progress before changing the application users experience.

Availability needs equally careful language. The current documentation marks learning from production as private beta, while the launch presents Continual Grading and Group Size 1 as research previews. A platform announcement should therefore be read alongside the status of its individual capabilities. Customers still need to establish which parts are accessible for their own deployment.

Trajectory's homepage also describes evaluation and approval workflows before model changes touch production. For a buyer, those controls make the learning loop a product responsibility that extends beyond starting a training run. Teams must preserve a usable relationship between the examples they supplied, the behavior they measured and the version they eventually approved.

Founders putting research into a usable product

An investor's founder record names Ronak Malde, Michael Elabd and Arjun Karanam. Ronak Malde is co-founder and CEO, according to his AI Engineer speaker biography, which also describes his work on Windsurf's SWE-1 and subsequent research at Google DeepMind. The three met at Stanford; Michael Elabd also worked at DeepMind, and Arjun Karanam has described his earlier Apple work in a Sequoia-hosted presentation.

The company's launch names Harvey, Rogo, Mercor and NVIDIA as organizations training on Trajectory. Harvey's own Sequoia-hosted discussion describes working with Trajectory on Nemotron training. Those relationships help locate the company in real application work without supplying undisclosed revenue, retention or paid-customer figures.

The surrounding market is asking related questions about who owns enterprise AI behavior. DevCuration's coverage of Applied Compute, Mintive and Cohere follows different parts of that responsibility, from specialized models to access controls and institutional infrastructure. Trajectory is concentrating on the customer's contribution to training. As its product reaches more teams, the lawyer's correction and the analyst's grading example become inputs someone must keep understanding, maintaining and carrying into the next model.