Trajectory bets $300 million valuation on AI models that learn on the job

Today’s AI models have a stubborn flaw. They make the same mistake on Monday that they made on Friday, unless someone retrains them.

Trajectory, a San Francisco AI infrastructure startup, has raised serious venture money on the premise that this flaw is fixable. It now has a self-service product meant to put that fix in the hands of ordinary businesses.

From stealth to $300 million in a few months

Trajectory came out of stealth in May 2026. Its founders are Ronak Malde, Michael Elabd and Arjun Karanam, part of a group of former researchers from Google DeepMind and Apple.

The company opened with a $15 million seed round, announced around May or June 2026 at a $115 million post-money valuation. That alone would be a respectable debut.

Then August arrived. Trajectory raised a $40 million Series A at a $300 million post-money valuation, which more than doubled its paper value in roughly a single summer.

Total disclosed funding now sits at approximately $55 to $65 million. The backers include Sequoia Capital, Nvidia and Bessemer Venture Partners.

On November 2, 2026, Trajectory launched a self-service platform. It lets businesses update their models from production data such as user edits and corrections, potentially as often as weekly.

The problem Trajectory says it solves

The limitation spans the industry. It applies whether a model comes from OpenAI, from Anthropic, or from an open-source project. Once a model ships, it does not improve from its errors without another round of training.

Companies working in specialized fields like law and finance usually handle this through fine-tuning. They take a general-purpose model and train it further on domain-specific material.

Trajectory’s pitch is automation. Instead of a research team manually curating data and retraining models, the platform takes the corrections users already make and feeds them back into the model.

The underlying technique is called self-distillation policy optimization. Stripped down, the model uses real-world edits and corrections to keep refining itself. Every human fix becomes a lesson instead of a wasted moment.

Who is already using it

Before the public launch, Trajectory worked with clients through closed beta partnerships. The names are notable in the AI application world.

Clay is an AI sales platform. Decagon builds customer support tools. Harvey is a legal AI firm.

The move to self-service marks a strategic shift. A closed beta with hand-picked partners is a controlled experiment. A self-serve platform is a bet that the product can work without the founders in the room.

What this means for the AI infrastructure race

The most direct threat is to the consulting model. If continuous fine-tuning can be automated, the case for expensive forward-deployed engineers weakens.

For investors, the valuation step-up tells its own story. Going from $115 million to $300 million within months signals strong appetite for infrastructure that sits between foundation models and end users.

There are real questions to watch, though. Learning from user corrections assumes those corrections are good. A model that absorbs every edit could also absorb every bad habit, and enterprises in regulated fields will want strong controls over what their models learn.