Build a SaaS control plane for post-training teams to define eval suites, compare model checkpoints, track reward signals, and recommend next post-training actions such as data mix changes or reward adjustments. The product would integrate with existing training infrastructure and produce experiment-level evidence linking eval movement to model changes.

Job postings show post-training is becoming a dedicated product and research function as labs shift toward reasoning-focused training methods like RLVR. Companies are moving from one-off fine-tuning toward continuous model improvement loops that need operational tooling.