Mirendil is a San Francisco-based AI research company building systems that automate AI research itself. Founded in early 2026 by former Anthropic researchers Behnam Neyshabur and Harsh Mehta — along with ex-xAI engineer Shayan Salehian and MIT graduate Tara Rezaei — the company trains frontier models end-to-end, covering pretraining, RL-based post-training, inference, and kernel optimization. Its core bet: that the next leap in AI comes not from better assistants, but from AI systems that can design experiments, search hyperparameters, evaluate models, and iterate autonomously — compressing months of research into days. The target customers are universities, pharma companies, and research institutions that need frontier AI R&D capabilities but lack the engineering headcount of an OpenAI or Anthropic. Mehta notably built Anthropic's first internal auto-research platform before leaving to productize that concept externally. A16z has called the approach "vibe research" — if it works, it could restructure who gets to run frontier-scale AI science.