Every AI tool you use today was trained on data up to a certain date, deployed, and then frozen. Your customer service bot that mishandled a refund request last Tuesday will mishandle the same request next Tuesday. The model has no idea it failed.
Trajectory, which announced its launch this week, is trying to change that. The startup raised $15 million at a $115 million valuation, led by venture firm Conviction, with backing from Bessemer Venture Partners and individual investors including Google DeepMind chief scientist Jeff Dean and AI research pioneer Fei-Fei Li.
The founding team comes from the labs that built the most capable AI systems in the world. CEO Ronak Malde was an AI researcher at coding tool Windsurf before joining Google DeepMind when it acquired that startup for $2.4 billion last year. Co-founder Arjun Karanam worked on Apple's Vision Pro headset. Michael Elabd came from Google DeepMind's robotics division.
The core product works like this. A company starts with an open-source AI model, rather than renting access to one from OpenAI or Anthropic. Trajectory then monitors how that model performs in the real world: when a customer support agent kicks a query to a human, when a user retries a prompt, when an output gets edited. Those moments of failure become training data. The model gets retrained, currently as often as once a week, and a new version ships automatically.
Trajectory already has paying customers. Legal AI firm Harvey is using it to build models that improve as lawyers interact with them. Sales tool Clay is doing the same. The challenge with spreading this beyond software tools is real: coding is easy to verify because code either runs or it doesn't. In customer service, legal drafting, or procurement, the definition of a good answer is murkier. Trajectory's platform helps companies define what success looks like so the model has something to learn from.
This sits inside a much larger story about AI deployment. An MIT study of 300 enterprise AI projects found that 95 percent showed no measurable effect on profit and loss. The issue was not the models. It was getting them to work inside real business environments. OpenAI and Anthropic just launched competing deployment ventures worth a combined $5.5 billion, both of them built around sending highly paid human engineers into companies to make AI tools actually work. Those engineers cost anywhere from $171,000 to $800,000 per year in total compensation, and that model scales badly: every new client needs more engineers.
Trajectory's argument is that a platform which lets a model learn from its own mistakes could, over time, reduce how much human intervention a company needs. Today that means weekly model updates. The founders say the goal is daily, then eventually per-interaction.
One thing to watch: the startup currently serves AI-native companies like Harvey and Clay. Selling this to a traditional manufacturer, insurer, or retailer is a different problem. Those organizations don't have the internal technical knowledge to define good training signals or approve new model versions. Trajectory will need to make that process close to automatic before the Fortune 500 becomes a realistic market.
The investors behind this round, Jeff Dean and Fei-Fei Li in particular, are not writing novelty checks. They understand the technical problem deeply and apparently believe the team has a credible path to solving it. That is worth noting.