AI Governance Platform

01Problem问题

AI-usage governance needs three things to hold at once: detection that works in English and Chinese, policy decisions that are deterministic, and an audit history that survives worker crashes and retries.AI 使用治理需要同时成立三件事：中英文都可用的检测、确定性的策略判定、以及在 worker 崩溃和重试后依然正确的审计历史。

02What I built我做了什么

• Policy engine as one pure function over the event and policy set, so every decision is reproducible.策略引擎是对事件与策略集的一个纯函数，每次判定都可复现。
• Separate governance and PII-detection APIs with role-based access; audit records keep categories, hashes and offsets, never the raw input text.治理 API 与 PII 检测 API 分开部署，按角色控制访问；审计记录只保留类别、哈希和位置，从不保存原始输入文本。
• Ingestion through a Postgres transactional outbox; workers claim events with SKIP LOCKED and record findings and incidents in a hash-chained history.通过 Postgres 事务性 outbox 接入；worker 用 SKIP LOCKED 领取事件，把发现与事件写入哈希链历史。
• Found and fixed a lease-expiry race: a stalled worker could finish an event another worker had already processed. It now locks the row and proceeds only while the status is still processing.发现并修复租约过期竞态：卡住的 worker 可能重复处理别人已完成的事件。现在会锁行，且仅在状态仍为 processing 时继续。
• Credentials use envelope encryption; 17 architecture decision records document each choice.凭据使用信封加密；17 份架构决策记录说明每个选择。

03Architecture架构

04Measured测量结果

05Limits, stated plainly局限，直说

• AWS infrastructure is designed in OpenTofu, not deployed. Throughput has only been measured locally.AWS 基础设施以 OpenTofu 设计，未部署；吞吐量仅在本地测量。