The Sequoia-backed company founded by former DeepMind and Apple researchers wants enterprises to continuously improve specialized AI models using feedback from real work.
Trajectory has raised $45 million in a Series A led by Sequoia Capital at a $300 million valuation, funding a platform that turns production feedback into continued model training.
The startup was founded by Stanford classmates Ronak Malde, Michael Elabd and Arjun Karanam. The team brings experience from organizations including Google DeepMind and Apple, where members worked on advanced models and infrastructure.
The Missing Feedback Loop
A company may deploy a capable general-purpose model but still find that the system struggles with the specialized standards required in legal work, finance or customer support. Employees can correct its mistakes, yet those corrections often remain isolated in an application rather than improving the underlying model.
Trajectory aims to automate the cycle of evaluating outputs, collecting feedback and post-training models. Customers can define what counts as a successful answer, examine failures and update specialized models without assembling a large internal research organization.
Its disclosed customers include Harvey, Rogo, Decagon, Mercor and Nvidia. These customers operate in demanding technical or professional settings where small improvements in quality and cost can matter.
The promise is not simply better benchmarks. It is a system that can improve as it encounters new cases, and where customers can adapt a model to the characteristics of their own workflows.
Competing With General-Purpose Models
Trajectory enters a rapidly developing post-training market. Enterprise customers must decide whether to build internal tuning capability, use vendors like Trajectory or rely on increasingly capable foundation model providers.
Continuous updates create their own risks. Training on production interactions can introduce regressions, unreliable feedback or unwanted changes in behavior if not rigorously evaluated.
The startup has not disclosed revenue or profitability. Its $300 million valuation reflects expectations that specialized models and post-training infrastructure will become an important layer of enterprise AI.
The next milestone is demonstrating that customers can repeatedly improve accuracy, cost and speed without creating more operational complexity than they remove.
Sources: TechCrunch · WIRED founder profile