# Refresh > Refresh builds artificial worlds where AI models are evaluated, trained, and improved through reinforcement learning. ## About We build realistic software environments that let AI models practice complex, high-value computer work — from EHR software to debugging production issues to enterprise workflows where success can be measured automatically. By turning real work into trainable environments with verifiable rewards, we help AI systems hill climb toward more capable, reliable behavior. ## Products & Research - PDFBench (https://refresh.dev/blog/pdf-bench): A benchmark for AI on PDF form-filling: 70 real-world tasks and 1,082 trials across frontier models. Claude Opus 4.5 led with a 71.1% overall pass rate, ahead of Gemini Pro 3 (61.7%) and GPT-5.2 (60.2%). - EMRBench (https://refresh.dev/blog/emr-bench): A benchmark for clinical computer-use agents on OpenEMR: 52 clinical tasks across 10 categories. Claude Sonnet 4.5 succeeded on 38 of 52 tasks versus 19 for Gemini 2.5 Computer Use. - Gauntlet (https://refresh.dev/blog/gauntlet-80k): An SFT dataset of 79,139 pytest-verifiable developer workflows drawn from 1,669 production Python repositories (37GB of source with git history); every example is verified against real tests. - Gauntlet 4K RLVR (https://refresh.dev/blog/gauntlet-4k-rlvr): A curated set of 4,000 reinforcement-learning tasks with verifiable rewards, distilled from Gauntlet and delivered in Harbor format for RLVR training. - Trajectories (https://refresh.dev/blog/trajectories-and-computer-1): A model-agnostic computer-use harness inside Harbor, plus a platform to upload Harbor jobs with one command, get a shareable URL, and inspect every step of every trial. - Bespoke environments (https://refresh.dev/one-pager): Custom, domain-specific simulation environments and datasets built together with domain experts in software engineering, finance, and healthcare. ## Key Pages - Homepage: https://refresh.dev - About: https://refresh.dev/about - Research and blog: https://refresh.dev/blog - Environments overview: https://refresh.dev/one-pager - Open roles: https://refresh.dev/careers ## Contact - Email: contact@refresh.dev - GitHub: https://github.com/refreshdotdev - Twitter: https://x.com/refresh_dev - LinkedIn: https://www.linkedin.com/company/with-refresh/