The Sequoia-backed startup wants companies to turn everyday user edits and corrections into weekly model upgrades
Trajectory, a startup built by former researchers from Google DeepMind, OpenAI and Apple, has opened a self-serve platform for post-training AI models on real production data.
The pitch is simple. Every time a user fixes an AI’s output, that correction becomes training material instead of wasted effort.
The platform went live on November 2, 2026. It moves Trajectory from a company that worked closely with a handful of partners to one that any qualifying business can sign up for and start using.
What Trajectory actually built
The mechanics run through a lightweight software development kit, or SDK. Companies plug it into their products, and it captures usage signals such as edits and corrections that users make to AI-generated output.
Those signals then feed post-training cycles. Post-training is the refinement stage that happens after a model’s initial training, where it gets tuned for specific behavior or tasks.
According to the company’s materials, those cycles can run on a weekly basis or more frequently.
Trajectory also offers a feature it calls continual grading, designed to scale human preferences. Put simply, it tries to turn the judgment of real users into a repeatable scoring system the model can learn from.
Then there is the GS1 algorithm, which the company says enables real-time learning from user actions.
Data handling is a central part of the sales pitch. Trajectory says client data usage adheres to SOC 2 standards, a widely used security and compliance framework, and it frames user control as a core operating principle.
The money and the customers
Trajectory was founded by Ronak Malde, Arjun Karanam and Michael Elabd. The company exited stealth mode in May 2026 before making this public, self-serve push.
The startup raised a $15 million seed round at a $115 million valuation.
Later rounds brought total investment to approximately $60-65 million. The standout was a Sequoia-led round that valued the company at $300 million.
The customer list includes Clay, Harvey, Decagon, Mercor and Rogo.
The research behind Trajectory’s approach suggests its machine learning solutions can outperform broader models in specific, narrowed tasks.
What this means for AI builders
There are real questions to watch. Training on production data means training on customer data, and that raises the stakes on privacy, consent and governance.
Trajectory’s emphasis on SOC 2 compliance and user control suggests it knows this is the first objection enterprise buyers will raise.
There is also a quality risk. User corrections are not always right, and a system that learns from every edit could absorb bad habits alongside good ones, which is presumably why continual grading sits at the center of the product.
