Trajectory raised $45 million in a Sequoia-led Series A for its AI training platform.
Why it matters: The financing backs tools designed to let lawyers, doctors, bankers and engineers improve AI systems using production data and expert feedback. The company is targeting organizations that want specialized models without building large machine-learning teams.
- Sequoia led the Series A, with participation from NVIDIA, Bessemer Venture Partners and Bain Capital Ventures.
- Trajectory was founded by Stanford classmates Ronak Malde, Michael Elabd and Arjun Karanam.
- The platform combines model training, evaluation and deployment using production data, user feedback and expert-defined standards.
- Trajectory's materials name Harvey, Rogo, Mercor, Decagon and NVIDIA; the materials do not specify whether they are paying customers, design partners or users.
San Francisco-based Trajectory announced the financing on November 2, 2026, in a company post. The company said Sequoia led the Series A, with participation from NVIDIA, Bessemer Venture Partners and Bain Capital Ventures.
Trajectory was founded by Stanford classmates Ronak Malde, Michael Elabd and Arjun Karanam. Malde and Elabd previously worked at Google DeepMind, while Karanam worked at Apple on generative foundation models and the Vision Pro, according to WIRED.
The company says its platform lets domain experts train, evaluate and deploy models using production data, user feedback and examples of a good result. It describes this as “continual learning,” meaning models are updated over time using feedback from real-world use. “Post-training” refers to improving an already trained model rather than building one from scratch.
Its research previews include Continual Grading, which applies user-defined standards across production data, and Group Size 1, intended to support training from individual production interactions. Human experts define and approve improvements. That distinguishes the approach from “recursive self-improvement,” in which an AI system would independently design and implement successive improvements to itself.
Trajectory says one customer trained more than eight models for complex, long-horizon “agentic tasks” in two weeks. In this context, agentic tasks are assignments in which a model must take multiple steps toward a goal. The claim appears in the company's materials and was not independently verified.
The company also reports that a Nemotron Lightning model improved from 0% to an 8.3% “all-pass rate” on Harvey's LAB benchmark. An all-pass rate measures the share of test cases in which the model satisfied every required condition. Trajectory's materials do not provide the benchmark's sample size, test conditions or comparison period. That result also was not independently verified.
Trajectory presents a public product page, while its website asks prospective users to join a waitlist. The company says it has raised $60 million across its seed and Series A rounds. WIRED reported a $15 million seed round in May 2026 at a $115 million post-money valuation.
By the numbers
- $45 million - Series A financing announced November 2, 2026
- $60 million - Combined seed and Series A financing reported by Trajectory
- 8.3% - Company-reported all-pass rate on Harvey's LAB benchmark, with sample size and test conditions undisclosed
Yes, but: Trajectory's performance claims, including the more-than-eight-models result and the 8.3% Harvey benchmark result, have not been independently verified. The company also has not specified the relationship of each named organization to its platform.