Trajectory has opened a self-service platform that lets organizations adapt models using their own tasks, scores and workplace feedback. Training, evaluation and deployment sit within one service. The company also describes two research previews aimed at learning users’ judging standards and training from production interactions, while customers already use the broader platform for specialist work.
Trajectory, a startup developing tools to adapt artificial intelligence models, has opened its post-training platform to the public. Post-training adjusts an existing model for a particular task. Users provide tasks and a way to score the results; the platform combines training, evaluation and deployment. Subject specialists supply the standards for the model’s work, with the service handling training infrastructure and orchestration.[1], [2]
Businesses can connect Trajectory to their applications and use feedback from completed work to retrain a model. TechCrunch reports that legal technology company Harvey and financial research platform Rogo already use the service to adapt models for specialized tasks. Those tasks include reviewing particular legal contracts and applying a chosen company-valuation method. The customer organization decides how a successful result should be judged.[2]
Two announced capabilities remain research previews. Continual Grading is designed to learn a user’s judging standards from examples and apply them to production data. Group Size 1 is intended to train from actual production interactions. Trajectory presents both as steps toward learning after deployment.[1]
Co-founder Michael Elabd told TechCrunch that the approach keeps people involved in training.[2]
