Trajectory Raises 45 Million Dollars From Sequoia and Nvidia to Let Nontechnical Experts Fine-Tune AI Models

Trajectory, a startup that wants to take model post-training out of the machine-learning lab and hand it to domain experts, has raised a 45 million dollar Series A led by Sequoia Capital at a 300 million dollar valuation. Nvidia, Bessemer Venture Partners and Bain Capital Ventures also joined the round.

The company's pitch targets a familiar bottleneck: the people who know where a model is wrong — a senior lawyer reviewing contract clauses, a credit analyst checking an underwriting draft — are rarely the people who can turn that judgment into training data. Trajectory's platform connects to business systems through a lightweight SDK and automatically captures how real users interact with model outputs: the edits they make, the errors they correct, the retries they trigger.

Those interactions become the raw material for continuous post-training. Companies can define optimization goals in plain terms, and the platform converts the captured feedback into model improvements for legal, finance and other vertical use cases, without requiring a machine-learning background.

The round fits a broader pattern: frontier-model gains are plateauing for generic chat, but enterprises keep paying for models tuned to their own workflows. Post-training — the layer between a released base model and a production system — has become one of the most contested infrastructure markets, attracting both dedicated startups and platform vendors.

Having Nvidia on the cap table is notable, since the chipmaker has been building its own post-training tooling around its hardware stack. A nontechnical feedback layer is complementary rather than competitive, but it signals how far the money is spreading around the fine-tuning workflow.

The open question for Trajectory and its peers is durability: feedback capture is easy to copy as a feature, and the hyperscalers or model labs themselves could bundle it. The bet Sequoia is making is that trust and workflow depth in regulated industries — where a wrong model output has real consequences — will be harder to displace than the SDK itself.