TensorTest (YC W26) — Evals, RL environments and training data for creative tasks

In their words

TensorTest (formerly Ritivel) builds evaluations, reinforcement-learning environments, and expert training data for AI labs developing creative agents, starting with video editing. An agent can produce a valid video file without delivering an edit that meets the client’s brief. Labs need reliable ways to measure creative quality, identify failures, and give agents feedback they can learn from. Our first benchmark, TimelineBench, gives agents client video-editing jobs with raw footage, audio, and a brief. It evaluates the finished work using a verifier calibrated on thousands of blind judgments from professional editors. Across 16 agents we tested, the best passed only 26.8% of jobs. Most failures came down to creative craft rather than broken output. We offer labs private evaluations to identify where their models fail, training environments that use our verifier’s scores as rewards, and expert-created training data, including professional editing demonstrations, preference pairs, and grading rubrics. These offerings connect evaluation to training: identify what an agent gets wrong, provide feedback on its attempts, and supply examples of better work. Our team comes from IIT Bombay and has trained language models and built agentic systems at Microsoft Research.

Founders, titles and bios as the company published them. Photographs are our own copies of the ones on that record, so they do not break when the original link expires.