Million

UniPat AI, a name derived from "Universe of Patterns," was founded in late 2025 as an infrastructure company building evaluation benchmarks and training data for AI agents and multimodal models. It designs real-world-grounded benchmarks covering the software-engineering ability of coding agents, the everyday web-operation ability of browser agents, and the visual reasoning ability of frontier multimodal models, with open-source projects including BabyVision, SWE-Vision, SaaS-Bench, EvoCodeBench, RoadmapBench, and Terminal-X, and sells the resulting synthetic training and evaluation data to other researchers and companies. It also builds in-house models such as UniScientist for autonomous scientific research and Echo, a prediction-intelligence system. Its founding team has backgrounds at Alibaba's Tongyi Lab, Moonshot AI, Tencent, and Peking University, with roughly 11 to 50 employees.

Founded in 2025 and based in San Francisco, idler is a Y Combinator Summer 2025 company that calls itself a frontier data research lab. It builds three things for the world's leading frontier labs: evaluation suites, programmatically generated datasets, and interactive reinforcement-learning environments used to measure and improve advanced models. Delivered domains span long-horizon software engineering, cybersecurity, recursive self-improvement, legal analysis, strategy and operations, and enterprise safety. Its first public benchmark, ShelfLife, is a digital twin of a multi-brand retail operation used for post-training evaluation. Team size is 12 to 14.

Comet (Comet ML, Inc.), founded in 2017 and headquartered in New York with an office in Tel Aviv, provides ML experiment tracking, model registry and production monitoring, and Opik, an open-source (Apache-2.0, self-hostable) LLM and agent observability and evaluation platform. Customers include Uber, Netflix, Autodesk, Etsy, AssemblyAI, Stability AI and Mobileye.

Robocurve is a Public Benefit Corporation building independent, reproducible benchmarks to measure how well frontier AI models can control real-world robots, positioning itself against the industry norm of curated demo videos. It has released Inspect Robots, an open-source evaluation framework, and runs standardized tests spanning tasks from block-stacking and sandwich-making to data-center construction, serving frontier labs, robot manufacturers and academic institutions. The founding team combines backgrounds in AI evaluation and robotics engineering; the company is part of Y Combinator's Summer 2026 batch.

BenchFlow, founded in September 2024 in San Francisco, is an AI agent evaluation lab that builds standardized, reproducible RL environments and benchmarks (SkillsBench, ClawsBench, PostTrain, FrontierPhysics) across coding, computer use, science, finance, healthcare and security, with an open-source sandboxed runtime and a community Benchmark Hub.

WisdomAI, founded in 2023 in San Mateo, CA, USA, develops tools that improve AI training data quality to reduce hallucinations in large language models. The company claims ConocoPhillips, Cisco, and Descope as early customers, and serves clients using major cloud data storage services including Snowflake, Google BigQuery, Amazon Redshift, Databricks, and Postgres.