English | 中文
AI Seed Project 是一个面向企业培训、AI 训练营和学习社群的飞书机器人培训评估系统：把群聊发言、图片、文件、测验卡片和互动行为转成学习行为证据，再通过 LLM 辅助评分、管理员审核、学习积分和排行榜看板追踪培训参与度。
English: a self-hosted TypeScript system for AI bootcamps, corporate training, and learning communities, with Feishu bot capture, LLM-assisted scoring, admin review cards, and a gamified learning dashboard.
Start here: Quick Start | Feishu setup | Admin guide | Student scoring rules
Search intent: for teams searching for 飞书机器人, 培训评估/企业培训系统, and 学习积分/排行榜/助教机器人.
Boundary: This is a self-hosted engineering template, not an official Feishu product. It requires a Feishu self-built app, Feishu group/Base configuration, and an OpenAI-compatible LLM provider; review consent, privacy, and data retention before using real learner data.
Companies invest massive budgets in training every year, yet evaluating training effectiveness remains an unsolved challenge.
| Pain Point | Current Situation | Consequence |
|---|---|---|
| Delayed Assessment | Statistics gathered only after training ends | Missed intervention window |
| Single-Dimensional | Only attendance and exams considered | Overlooking process performance |
| High Manual Cost | TAs record everything manually | Exhaustion and omissions |
| Learner Disengagement | Participants unaware of their performance | Lack of motivation to participate |
| Data Silos | Scattered across platforms | Cannot be consolidated or reused |
What we need is not another check-in tool, but an intelligent system that truly understands learning behavior.
A zero-friction, fully automated, AI-driven training evaluation system.
No app installation required, no new tools to learn---just communicate naturally in the Feishu group chat, and the system automatically captures everything, AI scores in real time, and the dashboard displays results instantly.
The system works like an indefatigable "super teaching assistant," silently active in the group 24/7. Every message, every image, every file, every interaction---all automatically captured, zero omissions.
Student posts a PPT screenshot
↓
System auto-identifies: this is a "learning outcome showcase"
↓
Classified under "Outcome Showcase" dimension, +8 points
↓
Leaderboard updates in real time ← entire process < 3 seconds
| Dimension | Code | Scoring Items | Description |
|---|---|---|---|
| Knowledge | K | K1 Daily Check-in, K2 In-class Quiz, K3 Knowledge Summary, K4 AI Correction | How much was learned and understood |
| Hands-on | H | H1 Assignment Submission, H2 Practice Screenshots, H3 Video Check-in | Whether they practiced and submitted work |
| Creativity | C | C1 Creative Sharing, C2 Like Interaction, C3 Deep Creation | Whether they showed creativity and new ideas |
| Social | S | S1 Group Message Interaction, S2 Peer Review Contribution | Whether they were active and helped others |
| Growth | G | G1 Video Learning, G2 Extracurricular Resource Sharing, G3 Sustained Activity | Whether they reflected, improved, and shared resources |
Semantic Scoring: Upgraded from keyword whitelists to LLM-based semantic understanding. When a message arrives, AI combines text, image understanding results, and file summaries to determine the contribution type, covering daily participation, AI-generated content, practice reviews, methodology sharing, resource recommendations, peer assistance, and formal tasks.
qwen3.5-flashpowers text scoring, image understanding, and TA Q&A, all via a unified OpenAI-compatible API- Images are asynchronously understood and cached first; subsequent scoring reads only the image descriptions to avoid blocking group chat replies with heavy image recognition
- Image-only messages support background re-scoring and restart recovery; real group message replay waits for async tasks to complete before closing the database
- AI Boot v3 supports
v3_shadowbypass mode: real database writes, auditable, replayable, but does not directly affect existing leaderboard scores
Proactive Praise Bot: The bot no longer passively waits for @ mentions. When it detects a student's outstanding contribution (total score ≥ 3), it proactively posts short, specific encouragement in the group. Production frequency can be controlled via toggle---defaulting to fewer but more precise, to avoid disrupting the group chat:
@Yang Bin This AI workflow nails the business pain points---the implementation feel is off the charts 🔥
@Wang Jing Effie The poster aesthetics and prompt approach are both on point---this creation is absolutely legendary 👏
Praise messaging adopts a "Xiaohongshu/Douyin comment section" style---short, specific, and internet-savvy, avoiding repetitive canned phrases like "great question" or "welcome other students."
Context-Aware Group Chat TA: When @-mentioned, the Singularity TA answers questions using recent group chat messages and files. Current scope: the 10 most recent chat messages + 2 most recent files, suitable for low-frequency training camp group chats; operational intents like score withdrawal, no-scoring, admin/review/adjustment are routed through dedicated handlers to avoid falling into casual conversation replies.
Weekly Leaderboard Settlement: Automatically posts the leaderboard report in the group every Thursday at 12:00---top 3 (🥇🥈🥉 with congratulatory messages) and bottom 3 (📌 with encouraging messages), powered by a systemd timer.
Hand-crafted CSS cyberpunk theme, designed to embed directly in Feishu group Tab pages.
Visual Style: Neon glow / Dark backgrounds / Data particle streams / Sci-fi HUD interface
Real-Time Leaderboard --- Grouped by tier, rank changes at a glance
Member Detail Page --- Badge wall + Radar chart + Five-dimension analysis + Growth timeline
| Level | Title | AQ Threshold | Meaning |
|---|---|---|---|
| 1 | 🌱 AI Prospect | 0+ | Just joined the camp, everything is possible |
| 2 | 🔬 AI Researcher | 32+ or meet Lv2 multi-path | Starting to think deeply, showing professional growth |
| 3 | 🎯 AI Operator | 64+ | Applying knowledge with significant results |
| 4 | 🧠 AI Advisor | 128+ | Influencing others, becoming a team thought leader |
| 5 | ⚡ AI Singularity Player | 224+ | Excellence across all dimensions, pushing cognitive boundaries |
Level-ups no longer depend on settlement cycles---promotions happen continuously once requirements are met. Lv2 additionally supports multi-path evaluation: 24+ points with C/S/G signals, 32+ points on the strong practice path, or 20+ points on the multi-dimensional activity path can all trigger promotion, preventing "hollow promotions" built solely from check-ins and video points.
Trainers don't need any technical knowledge. All management operations are performed via group chat keywords:
admin / admin panel / control panel → Admin control panel (open camp, open window, graduation settlement)
quiz / in-class quiz / exam → Draw questions from Feishu Base question bank, send interactive quiz cards
peer review / peer vote / vote → Peer review voting card, auto-calculate social scores
board / ranking / leaderboard → Ranking ladder card, one-click jump to Web Dashboard
adjust / manual adjustment → Manual score adjustment card (bypass auto-scoring limits)
member / member management → Member role management card (hide ranking / change role)
The question bank is managed via Feishu Base spreadsheets---add, delete, edit, and query directly in the spreadsheet, and the system syncs automatically.
- Per-camp caps --- Each scoring item has an independent score cap per camp to prevent score farming
- Idempotent deduplication --- The same message will not be scored multiple times
- AI quality verification --- Scoring must include contribution type, evidence, confidence level, and reasoning
- Admin review queue --- AI scoring results push review cards for one-click approve/reject
- Shadow safety valve --- v3 scores can be written to the database for observation without directly affecting the leaderboard
- Operational correction intents --- Messages like "don't add points / withdraw points / just casual chat" won't enter open conversation
- Manual score adjustment --- Admins can bypass auto-caps for manual corrections
| Highlight | Description | |
|---|---|---|
| 🎯 | Zero Friction | No tools to learn---group chat is the assessment |
| 🤖 | AI-Driven | Multi-modal LLM understanding, beyond the keyword era |
| ⚡ | Real-Time Feedback | Scored and on the leaderboard within 3 seconds, instant positive reinforcement |
| 🎮 | Gamified | Five-level growth system + badges + leaderboard |
| 📊 | Multi-Dimensional | 5 dimensions, 15 metrics---comprehensive profiling |
| 🔧 | Zero-Code | Group chat commands + spreadsheet management---trainers are admins |
| 🛡️ | Anti-Cheat | Rate limiting + deduplication + AI quality checks---ensuring fairness |
| 📦 | Lightweight Deploy | Single-server SQLite + systemd, live in 10 minutes |
| Scenario | Description |
|---|---|
| Corporate AI Training Camp | Zero TA overhead, fully automated, real-time data visualization |
| New Employee Onboarding | Auto-track participation depth, generate personal growth reports |
| Product Manager Bootcamp | Five-dimension model maps to product competency profiles, radar chart at a glance |
| Book Club / Learning Community | Gamified level system for sustained engagement, maintaining long-term activity |
| Hackathon | Real-time dashboard + team dimension scoring, maximizing competition atmosphere |
The system uses a modular architecture, with the following directions offering low-cost extensibility:
| Data Layer | IM Adapter Layer | AI Capability Enhancement |
|---|---|---|
| Multi-group linkage | WeCom adapter | AI personalized learning suggestions |
| Cross-camp data comparison | DingTalk adapter | Adaptive difficulty question bank |
| Data export reports | Webhook integration | Smart group matching |
| Student profile accumulation | Mobile H5 adaptation | Training ROI quantitative analysis |
Frontend: React 18 + TypeScript + Vite (hand-crafted CSS, cyberpunk theme)
Backend: Fastify + TypeScript (Node.js)
Database: SQLite (better-sqlite3, WAL mode)
IM: Feishu Open Platform SDK (WebSocket long connection)
AI: Qwen3.5 Flash / OpenAI-compatible LLM (text scoring, vision understanding, TA Q&A)
Schedule: systemd timer (leaderboard settlement every Thursday at 12:00)
Deploy: systemd + Nginx + Let's Encrypt, runs on a single server
git clone https://github.com/Ethan-YoungQ/ai-seed-project.git
cd ai-seed-project
npm installcp .env.example .envEdit the .env file and fill in:
| Variable | Description |
|---|---|
FEISHU_APP_ID |
Feishu custom app App ID |
FEISHU_APP_SECRET |
Feishu custom app App Secret |
FEISHU_BOT_CHAT_ID |
Chat ID of the group where the bot resides |
LLM_PROVIDER |
LLM provider (recommended: aliyun, compatible with OpenAI-style Chat API) |
LLM_API_KEY |
LLM API Key |
LLM_TEXT_MODEL |
Text scoring / TA Q&A model (production recommended: qwen3.5-flash) |
LLM_VISION_MODEL |
Image understanding model (production recommended: qwen3.5-flash) |
AI_BOOT_ENGINE_MODE |
AI Boot engine mode: legacy / v3_shadow / v3_live |
See .env.example for complete environment variable documentation.
- Go to Feishu Open Platform to create a custom app
- Enable the Bot capability
- Add permissions:
im:message:receive,im:chat,im:message:send - Set up event subscription:
im.message.receive_v1 - Select Long Connection mode (recommended)
See docs/feishu-setup.md for detailed steps.
# Development
npm run dev
# Production build
npm run build
node dist/main.js
# Run tests
npm test
# Replay existing Feishu group messages in dry-run mode
npm run ai-boot:replay-feishu-messages -- \
--messages /tmp/feishu-messages.json \
--database-url ./data/app.db \
--camp-id default \
--chat-id oc_xxxOpen http://localhost:3000/dashboard in your browser
ai-seed-project/
├── src/
│ ├── domain/v2/ # Scoring domain model (ingestor, settler, levels, promotion-announcer)
│ ├── routes/v2/ # API routes (board, ranking, member detail)
│ ├── services/feishu/ # Feishu integration (bot, cards, message handling, AI Boot v3, promotion announcement)
│ ├── storage/ # SQLite repository
│ └── config/ # Configuration and defaults
├── apps/dashboard/ # React Dashboard (Vite)
│ ├── src/components/ # UI components (cyberpunk theme)
│ ├── src/hooks/ # Data hooks (useRanking, useMemberDetail)
│ ├── src/lib/ # Utility functions (badge-engine, colors, api)
│ └── src/routes/ # Page routes
├── tests/ # Vitest test suite
├── scripts/ # Operations and deployment scripts
│ ├── ops/deploy-app.sh # Application deployment
│ ├── ops/deploy-timer.sh # Timer deployment
│ └── ops/backup-db.sh # Database backup
├── deploy/systemd/ # systemd units
│ ├── ai-seed-project.service # Main service
│ ├── weekly-ranking.service # Ranking script service
│ └── weekly-ranking.timer # Timer for every Thursday at 12:00
├── docs/
│ ├── admin-guide.md # Admin guide (no technical background required)
│ ├── student-rules-guide.md # Student scoring/promotion/badge rules
│ ├── feishu-setup.md # Feishu app configuration steps
│ └── project-pitch.md # Project introduction and use cases
└── .env.example # Complete environment variable template
- Node.js 18+
- 1 Core / 1 GB RAM
- Feishu Enterprise Edition (Open Platform permissions)
- Alibaba Cloud Bailian or other OpenAI-compatible LLM API account
npm install
npm run build
# Using systemd (recommended)
sudo cp deploy/ai-seed-project.service /etc/systemd/system/
sudo systemctl enable ai-seed-project
sudo systemctl start ai-seed-project
# Or run directly
node dist/main.js| Method | Path | Description |
|---|---|---|
| GET | /api/health |
Health check |
| GET | /api/v2/board/ranking |
Leaderboard data |
| GET | /api/v2/board/member/:id |
Member details |
| GET | /dashboard |
Dashboard SPA |
This project is designed as a reusable template:
- Swap IM platform: Modify
src/services/feishu/to adapt to other IM platforms (WeCom, DingTalk, etc.) - Swap AI engine: Modify
src/services/llm/provider-config.tsandsrc/services/v2/llm-scoring-client.tsto integrate other OpenAI-compatible LLMs - Customize scoring dimensions: Modify the scoring rules and dimension definitions in
src/domain/v2/ - Customize dashboard theme: Modify the CSS variables in
apps/dashboard/src/
| Doc | Description |
|---|---|
| Admin Guide | Admin guide (no technical background required) |
| Student Rules | Student scoring/promotion/badge rules documentation |
| Feishu Setup | Detailed Feishu app configuration steps |
| Project Pitch | Project introduction and use cases |
| .env.example | Complete environment variable documentation |
Issues and Pull Requests are welcome!
- Fork this repository
- Create a feature branch (
git checkout -b feature/amazing-feature) - Commit your changes (
git commit -m 'feat: add amazing feature') - Push the branch (
git push origin feature/amazing-feature) - Open a Pull Request
Let training be more than a one-person show.
Let AI see everyone's effort. Let data tell every growth story.
AI Training Camp Evaluation System Powered by Feishu Bot + Qwen3.5 Flash + Gamification