Trajectory raises $15M from Conviction to build AI agents that continuously learn on the job

Trajectory raises $15M seed from Conviction

Trajectory, a three-person founding team out of San Francisco, has raised a $15M seed round led by Conviction to build what they describe as a platform for continual learning — AI agents that get measurably better every day rather than resetting to baseline with each deployment.

The founding team

Ronak Malde was an AI researcher at Windsurf before its $2B acquisition by Google DeepMind, which he joined briefly before leaving to start Trajectory. Michael comes from foundation model research at DeepMind, with a background in robotics. Arjun spent time at Apple on Vision Pro and multimodal foundation models.

“We raised $15,000,000 from Conviction. We're building the platform for continual learning. We're working with some awesome companies — Harvey, Decagon, Clay, Rogo — basically building agents that learn online. We have models that are 10 times smaller than the frontier models that are able to beat them. But also we're improving 1% every day.”

The core argument

Every major lab is focused on making models smarter in the abstract — building what Michael calls "a smart PhD student." The problem is that student always shows up on day one, making the same mistakes every time. Trajectory's bet is that domain-specific experience matters more than frontier capability for most production use cases. Their models are trained to understand a specific company's workflows, primitives, and reward signals through lightweight post-training on live usage data — edits, retries, corrections — rather than expensive frontier-scale compute.

The result, according to Michael, is models 10x smaller than frontier models that outperform them on task — and more importantly, models that improve 1% per day through continuous learning.

Customer mix and commercial logic

Current customers include Harvey, Decagon, Clay, Rogo, and Recore — all application-layer AI companies building verticalized products. Arjun's read on what draws them is that these companies have already made a bet that product craft and domain expertise will hold as a competitive advantage. Right now their only tool for working that expertise into a model is prompt engineering. Trajectory positions itself as a research lab in their back pocket, giving them the ability to directly modify models and inference harnesses rather than relying on prompts alone.

The $15M and capital-light approach signal that Trajectory is not trying to compete on foundation model scale. The play is post-training on customer usage data — a narrower technical problem with a clearer path to production value and, presumably, a path toward making its customers less dependent on upstream model providers as their proprietary data compounds over time.

Trajectory is hiring in San Francisco only.