Until now, shaping a model took a frontier lab. Starting today, anyone who knows what “good” looks like can train a model to deliver it.
Today we're launching the Trajectory platform, the easiest and most intuitive platform for post-training on production data, and announcing our $45M Series A led by Sequoia at a $300M valuation, with participation from NVIDIA, Bessemer, and Bain Capital Ventures.
For the last few years, a handful of research labs decided how AI behaves. They picked the data, wrote the rewards and chose what "good" looked like. Everyone else rented the result.
That made sense when training a model required a research team, a GPU cluster and years of accumulated know-how. But the people with subject matter expertise (the lawyer drafting a merger agreement, the banker building a model, the engineer shipping production code) aren’t in the lab. Their judgment is the most valuable training signal there is, and so far it's gone unused.
Every company building on AI is facing climbing token bills, and models still miss critical context. Post-training can help, but getting started often takes months of R&D and complex infrastructure.
That's about to change. The best AI won't only come from researchers in a lab. It will come from the experts closest to the work. For the first time, the people who know what “good” looks like can own the model.
What we're launching
The Trajectory platform gives every company the power of a full Reinforcement Learning (RL) research team.
Picture a world where every company has its own intelligence. A law firm's model gets sharper with every contract its lawyers review. A bank's model learns its analysts' standards. A coding tool's model learns its users' codebases. Expertise compounds instead of staying locked in people's heads.
With our platform, you bring the one thing nobody else can supply: your value function, meaning your expertise. Trajectory handles the rest.
It's the first self-serve product for post-training on production data, so you don't even need to call us to start. Plug in a set of tasks with a way to score them, and you can train, evaluate and deploy your own model on one platform. What used to mean stitching together four products (sandboxes, evals, training and deployment) is now one product and about five lines of code.
The platform lets you:
- Train and deploy your own model. One click to train, one click to deploy. All without managing sandboxes, orchestration or infrastructure.
- Teach it what good looks like. Continual Grading (research preview) learns from a few examples of your judgment and applies those standards across your production data, reducing the need for manual labeling.
- Let it learn on the job. Group Size 1 (research preview) lets you train on real production interactions, so your model can keep improving after it ships, without months spent building training environments.
These two research previews are the first of many releases towards continual learning: AI that improves from the work it does, the way people do.
Builders at Harvey, Rogo, Mercor and NVIDIA are already training on Trajectory. One customer trained more than eight models for complex, long-horizon agentic tasks in just two weeks.
Working with NVIDIA and Harvey, we post-trained Nemotron 3 Ultra on frontier legal tasks within 24 hours of its release. We also took a Nemotron Lightning model from 0% to an 8.3% all-pass rate on Harvey’s LAB benchmark, matching Opus 5.5’s score at 100x lower cost and 3.8x faster speed.
Making this possible takes both frontier research and thoughtful product design. Our research team contributed to Gemini 3 at DeepMind, GPT-4o at OpenAI, and Muse infrastructure at Meta Superintelligence, while our product team built and launched unique interfaces at Stripe, Netflix, and Apple. Our strength is where those two meet: frontier research, shipped as a product anyone can use.
The people closest to the work know what a better model looks like. Trajectory gives them the tools to build it. Start training your own model today.
If you want to help build that future, we’re hiring.