English | 简体中文
Five local-first, bring-your-own-key AI workbenches for real vertical work — law, teaching, e-commerce ops, sales, and content creation. One repo, five self-contained apps, zero dependencies.
-2e5d43.svg){kind=icon}
Generic chatbots answer questions; vertical work needs workbenches — tools that hold your data, run real domain mechanisms locally, and use AI only where it adds value, grounded and validated. Each app here follows the same contract:
- Local-first: a single
python server.pyon127.0.0.1, your data never leaves your machine. - BYOK: bring your own API key (OpenAI / Anthropic / Gemini / DeepSeek / Kimi / GLM / any OpenAI-compatible relay). Keys live in your browser, travel per-request, are never stored or logged server-side.
- Offline mock mode: every app is fully usable with no key at all — deterministic rule/template paths stand in for the model.
- Grounded AI: model output is validated against your data (verbatim-quote grounding, number cross-checks, coverage validators); ungrounded output is demoted or stripped, never silently trusted.
- Real artifact out: every app exports a downloadable draft document (report / 教案 / 简报 / review / content pack) — a decision record you edit and own.
- Zero dependencies: Python 3.8+ standard library only. No pip install, no node, no docker.
| # | App | For | Core mechanisms | Tests |
|---|---|---|---|---|
| 1 | CaseAnchor 类案检索工作台 | Lawyers who must produce similar-case retrieval reports (类案检索报告) | Judgment structure parser · BM25 hybrid retrieval · 5-dimension similarity with counterfactual sensitivity · grounded 裁判要点 extraction with rule/AI cross-check · report draft with 同案不同判 divergence analysis | 63 |
| 2 | LessonSmith 备课工坊 | K-12 teachers producing 教案 and tiered exercises under the 双减 mandate | Objective parser with 学段-verb alignment · curriculum-standard KB · constraint-based lesson scheduler with what-if replan · validated per-segment generation · three-tier exercise engine + 教案 export | 75 |
| 3 | ShopPulse 店铺体检台 | E-commerce operators explaining "why did GMV move" every Monday | CSV ingest with platform header presets + GB18030 · LMDI factor decomposition (visitors × conversion × AOV) · weekday-seasonal MAD anomaly detection · what-if target simulation with reverse-solve · playbook diagnosis + grounded AI 诊断简报 | 88 |
| 4 | FollowLoop 跟进台账 | Sales reps juggling dozens of open opportunities | Customer/interaction ledger · stage×intent cadence engine with promise override · explainable win-probability score with per-factor counterfactuals · grounded 话术 drafting validated against your notes · pipeline review export | 110 |
| 5 | DraftDeck 选题工坊 | Creators turning one idea into platform-ready content | Angle matrix + title formula bank · explainable title scoring with clickbait penalty · four-role AI pipeline (策划→标题→撰稿→审读) with typed contracts and a bounded revise loop · deterministic platform adapters (公众号/知乎/小红书/抖音口播) with rule traces · content pack export | 53 |
Each subdirectory is a complete standalone product with its own bilingual README, tests, fixtures, sample data, and design system — five deliberately distinct visual identities, no shared framework.
git clone <this-repo>
cd <repo>/caseanchor # or lessonsmith / shoppulse / followloop / draftdeck
python server.py # Python 3.8+, nothing to install
# open the printed http://127.0.0.1:80XX addressEvery app ships a one-click sample dataset — click the sample button on the empty screen and the full pipeline runs offline in mock mode.
Open the AI 设置 dialog in any app: pick a provider preset or paste any OpenAI-compatible base URL (relay/中转站, LM Studio, vLLM), paste your key, hit test. The key is stored in your browser's localStorage only, masked, clearable, sent per-request through your local backend directly to the provider you chose — never persisted or logged server-side. Long AI tasks run in the background with progress, cancel, refresh-resume, history, and per-task token usage.
Every app in this repo passed the same machine-checked gates before inclusion:
- Full test suite green (389 tests total across the five apps).
- A 24-cell visual-integrity audit (4 viewports × 6 UI states — empty/loaded/loading/complete/error/modal) with zero blocking issues, machine-verified geometry (no occlusion, no overflow, reachable modals).
- Claim-checked READMEs: every capability claim maps to code and tests; limitations stated honestly (each app's README says exactly what is heuristic, what is lexical, what is a draft).
- Secret scan clean; keys never logged.
.
├── caseanchor/ # 1 · legal similar-case retrieval workbench (port 8056)
├── lessonsmith/ # 2 · teacher lesson-prep workshop (port 8057)
├── shoppulse/ # 3 · shop metrics examiner (port 8058)
├── followloop/ # 4 · sales follow-up ledger (port 8059)
├── draftdeck/ # 5 · content creation workbench (port 8060)
├── LICENSE # MIT, covers the whole suite
└── README.md # this file
These are honest tools, not magic: retrieval is lexical+structural (not neural search), scores are explainable heuristics (not trained models), AI output is grounded and validated but still needs human review, and every exported document is a draft for a professional to own. Each app's own README carries a specific limitations section — read it before relying on the output.
MIT — the whole suite.