📄 中文摘要
SAFi 是一个开源项目，旨在为大型语言模型（LLM）提供一种新的反馈机制。与传统的守卫系统不同，SAFi 不仅在请求时进行检查，还能记住 AI 的行为，检测其是否偏离预期的角色，并通过数学方法将其引导回正轨。该系统由一系列专门的模块（称为“学院”）组成，每个模块负责特定的任务。SAFi 已在生产环境中部署，处理了超过 1600 次经过审计的交互，展示了其在实时指导 LLM 方面的有效性。
我构建了一个在运行时使用 NumPy 指导 LLM 的反馈循环
出处: I Built a Feedback Loop That Coaches LLMs at Runtime Using NumPy
发布: 2026年2月12日
📄 English Summary
I Built a Feedback Loop That Coaches LLMs at Runtime Using NumPy
SAFi is an open-source project designed to provide a novel feedback mechanism for large language models (LLMs). Unlike traditional guardrail systems that merely check requests at the door, SAFi remembers the AI's behavior, detects when it drifts from its intended character, and mathematically coaches it back on course. The system consists of a pipeline of specialized modules, referred to as 'faculties,' each responsible for specific tasks. Deployed in production, SAFi has handled over 1,600 audited interactions, demonstrating its effectiveness in real-time coaching of LLMs.