AI 应用与工作流开发。把一个具体问题,推进到能运行、能验证、能交接。
我是恩宇,现居郑州,自 2026 年 6 月参与企业 GEO 与 AI 应用实践。我用 Python 做知识检索、AI 接口集成和内容工作流,关注资料的出处、请求的流转、失败的处理,以及交付后能否由别人继续运行。
作品网站 · 操作三个重点作品 · 简历 · 合作方式与交付范围 · 直接联系
| 你的问题 | 我能承担的实现 | 可检查的交付 |
|---|---|---|
| 资料分散,回答难以追溯 | 整理文档结构,接入本地检索,处理版本、权限与引用 | CLI/HTTP、资料与权限说明、固定回归、备份恢复步骤 |
| AI 接口偶发失败,缺少排查线索 | 集成 HTTP/SSE,区分超时、取消、协议与容量问题 | 本地网关、元数据报告、复现脚本、故障恢复说明 |
| 内容制作依赖手动串流程 | 串接素材清单、时间轴、字幕、FFmpeg 和媒体检查 | 可重跑任务、校验清单、交付包、人工审片入口 |
| 改过 AI 工作流,却说不清有没有改善 | 建立用例/响应记录、执行机械约束、对比失败与回归 | 版本报告、可复跑命令、待人工确认的问题 |
适合从范围明确的原型、内部工具或现有流程改造开始。先对齐问题、可用资料和验收目标,再交付最小可运行版本与交接记录。涉及真实模型、生产身份或客户数据时,需要在具体项目中完成相应验证。
你可以带着现有流程、脱敏样例和一个验收目标来聊。我们先确认输入、输出和依赖,再拆解实现范围;交付包含运行命令、测试记录、已知限制与交接说明。详见合作方式。
| 工程 | 我实现的部分 | 直接查看 |
|---|---|---|
| Enterprise RAG Starter | 三格式文档入库、版本/生效期、身份过滤、引用与拒答、SQLite 备份恢复 | 源码 · 运维流程 · 在线操作 |
| RelayAudit / AgentGuard | 固定上游代理、HTTP/SSE 保真、独立观察队列、容量控制、异常报告 | 源码 · 容量与故障 · 报告检查器 |
| AI Video Workflow Engine | 素材/时间轴/字幕、真实渲染与 QA、可校验交付包、人工审片门禁 | 源码 · 交付包 · 54 秒原创动画 |
公开网页 RAG 为合成资料上的词法检索;网关报告不确认真实模型身份;动画样片的人工批准只适用于其确切文件。工程证据与适用范围在各仓库相邻说明。
本轮工程审查 · 全部项目与阶段 · 项目贡献记录 · 多租户 HTTP 工程 · 响应评测器
复现命令、评测边界与文件证据
- RAG: local architecture / trust boundary, existing lexical UI screenshot, and local CLI / HTTP QuickStart.
python local_service.py initstarts a new local index; follow the QuickStart for fictional sources and private token files. The browser preview is lexical and does not demonstrate authenticated local-service access. The pre-repair45e2daa8baseline passed 48 Python + 10 Node tests and 4 lexical / 9 answerability cases; current excerpt-fix results are recorded in the repository verification section. The recorded synthetic_unverified local candidate run is 60/62 over 56 distinct questions (required lexical-contract cases 60/60), with two semantic-paraphrase misses preserved and zero real-model calls. The user confirmed completing review of all 62 cases; no per-case adjudication or expected-answer edits are inferred. Provenance remains synthetic_unverified, not an independently verified Golden Set; enterprise and semantic quality are unverified. - RelayAudit:
Agent → fixed-route proxy → configured upstream, with independent bounded parser/detector queues and separate notifications. Protocol / source · five-minute synthetic QuickStart. Runpython -m relay_audit guard demo --out agentguard-demo/profile-001in a fresh output directory. Original Actions evidence is retained; its downloadable artifact has finite retention and is not the stable demo entry. C1/N100 and C8/N200 measurements retain p50/p95, throughput and Python peak allocations; zero extra model requests does not mean zero latency or 100% performance. - Video: architecture · nine-shot contact sheet · sample commands and rights manifest. Generate via
python examples/original_54s/generate.py --output output/original54 --font /path/to/licensed-font.ttc, then use the linked FFmpeg workflow command. The 2026-10-30 source CI passed 56 tests, including delivery-bundle checks. The exact 54-second, nine-shot original animation is user-viewed APPROVED; artifact-specific review pins SHA2560a7f73e9a5a014a352aa5f27a2852b39e558546ccc38f93016785cf2fba0e1a9. It is animation, not real-person talking-head footage or customer/model/commercial acceptance.
Each repository records mechanical tests separately from AI evaluation and human acceptance. These are personal tools, prototypes and demos; no customer production, commercial success or model-identity accuracy is implied.
The portfolio index retains the original sixteen public project sources, grouped as other flagships, supporting work, templates and experiments. It also labels legacy prototypes and upstream-derived work. Private source archives and their component directories are not counted as production products.
dbskill is third-party work from dontbesilent, under CC BY-NC 4.0 / non-commercial terms. I do not present it as an original project or change its license. Company/client material, GEO work under IP review and private voices/data stay outside original Selected Work.
I am interested in AI application engineering, retrieval, evaluation, automation and content workflows. The practical standard is a runnable path, evidence matched to risk, explicit failure states and a handoff another person can reproduce.
Personal website · X · GitHub
Telegram @ethanyu520 · Email: zhouey314@gmail.com · WeChat / 微信:Ethan 恩宇
微信扫码添加 Ethan 恩宇。
可以直接带着一个具体问题来聊:现在哪一步最费时间、有哪些资料和系统、怎样算做成。我会据此拆解可交付范围与验收方式。先看合作方式与交付范围。
