Trajectory.ai cut legal AI model onboarding from three months to one week
Ronak Malde describes how Trajectory.ai's first Harvey engagement took three months to stand up. The most recent took a week, driven by a training system built on a technique rarely attempted outside academia.
Trajectory.ai’s Ronak Malde describes a sharp compression in how long it takes to bring a new legal AI customer onto the platform. The first engagement with Harvey required roughly three months. A follow-on engagement came in under a month. The most recent new customer was trained and running within a week.
The first engagement, it took like three months to set up the entire thing. We're kind of building the airplane as it was flying and now we're able to train a model with Harvey in under a month. And then onboarded a new customer and within a week we were able to train a model and really get the flywheel going. Ronak Malde
The speed gain rests on Trajectory.ai’s Continuous Lara system, which runs training jobs concurrently rather than sequentially. Malde says eight concurrent runs finish in roughly the same wall-clock time that three sequential jobs would require, a parallelism gain that compounds as the customer pipeline grows. Under the hood, the system uses self-distillation policy optimization (SDPO), a technique Malde says has been demonstrated in academic settings but that no one has previously scaled to production use cases.
The urgency around that last mile matters in legal contexts specifically. Malde puts it plainly: “getting 80% of the way there is the same thing as zero,” because any remaining gap still demands full expert involvement. That framing also shapes how Trajectory.ai collects feedback signals. Binary thumbs-up or accept/reject inputs, standard in most AI coding tools, are too noisy to rely on. Malde’s observation is that users rarely press those buttons at all, which means the training signal that matters has to come from somewhere else.