REX AI is a multi-agent training operations MVP developed for Andalusia Academy. It helps staff turn verified course information into teaching plans, training documents, and marketing campaigns. Learners get course guidance from the same published catalog, while staff can research training tenders in Egypt and Saudi Arabia.
The main engineering decision: a model can suggest facts, draft content, and propose actions. The application verifies record state and source provenance, and staff make the decisions that publish or approve work.
Quick start · Product preview · Architecture · Verification
| Curriculum Studio | Campaign Studio |
|---|---|
 |
 |
Screenshots are from the isolated demo workspace using fictional data. Labels and navigation can differ from the latest build.
See a generated training proposal
| Workspace | What it delivers |
|---|---|
| Learning Advisor | Arabic and English guidance, a guided qualification flow, source-cited course answers, constraint-aware recommendations, and consented lead capture. |
| RFP Research | JSON, CSV and Excel imports; LangSearch MCP discovery and public tender directories; Egypt/Saudi Arabia filters, expiry checks, deduplication and dashboard alerts. Search leads require verification. |
| Curriculum Studio | Structured modules, hours, activities, assessments and capstones, with revision history and PDF export. Staff can compare plans against verified framework criteria. |
| Document Studio | Ten template categories, Arabic and English output, branded PDF and editable Word files, source references, and versioned review. |
| Campaign Studio | Campaign strategy and variations for social captions, Facebook/Instagram, LinkedIn, email, WhatsApp, headlines and calls to action. Approved export is available after review; distribution is manual. |
| Instructor workspace | Persistent staff conversations and a planner that proposes actions across the workflows. It uses allowlisted internal APIs, requires confirmation for writes, and keeps approval decisions separate. |
flowchart LR
L["Learner · Course library and advisor"] --> UI["Streamlit"]
S["Staff · Review pages and Instructor chat"] --> UI
UI --> API["FastAPI · validation and staff access"]
API --> A["Domain services · catalog and five agents"]
API --> O["Instructor planner · 44 allowlisted tools"]
O --> API
A --> DB[("SQLite · records, review history and FTS5")]
A --> F[("Local files · originals, covers and exports")]
A --> AI["Groq / Gemini · structured generation"]
A --> X["LangSearch MCP · RFP discovery"]
A --> P["Puter · optional course covers"]
M["Local stdio FastMCP · 10 tools"] --> A
The default retrieval path is SQLite FTS5 over staff-verified course facts. Gemini embeddings and embedded Chroma are optional. The Instructor planner uses internal API calls; the local MCP server is a separate integration for external clients.
flowchart TD
U["Upload original PDF, DOCX or Markdown"] --> R{"Staff source review"}
R -->|Reject with reason| J["Preserved source and audit event"]
R -->|Verify facts and publish| C["Published course"]
C --> L["Learner answers and recommendations"]
C --> D["Curriculum draft"]
D --> A{"Staff reviews curriculum"}
A -->|Revise| D
A -->|Approve| P["Approved teaching plan"]
P --> G["Generate training documents"]
P --> M["Draft marketing campaign"]
G --> E["Review and download PDF / DOCX"]
M --> F["Review and export campaign"]
Engineering highlights
- Traceable data: original files retain their format and checksum. Rejected uploads cannot feed publication. Retrieval checks the current published record and source revision.
- Reviewable changes: versioned outputs and audit history preserve decisions. Withdrawn or superseded sources can block dependent approvals or exports.
- Controlled orchestration: request identifiers protect against duplicate execution. Decision previews are rechecked before execution. A new curriculum must be reviewed before dependent work can use it.
- Bounded inputs: upload and Office archive limits, isolated demo storage, and source validation reduce accidental exposure and unsafe processing.
- Optional integrations: provider errors surface without fabricating output. Course publication does not depend on Puter; cover generation runs once per revision after publication, and browsing reads the stored image.
Use Python 3.14 and run the commands from the repository root. In PowerShell:
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements-dev.txt
if (-not (Test-Path .env)) { Copy-Item .env.example .env }Set the following in the ignored local .env. Replace the placeholders locally.
REX_APP_MODE=academy
REX_CATALOG_ADMIN_KEY=your_local_staff_secret
REX_AI_PROVIDER=groq
GROQ_API_KEY=your_local_provider_key
REX_RETRIEVAL_BACKEND=sqliteStart FastAPI and Streamlit in separate terminals:
# Terminal 1
.\.venv\Scripts\python.exe -m uvicorn backend.app.main:app --host 127.0.0.1 --port 8000# Terminal 2
.\.venv\Scripts\python.exe -m streamlit run frontend/app.py --server.port 8501Open REX AI or the interactive API reference. The Academy catalog starts empty. Staff can sign in, upload a verified course source, check its details, and publish it. AI drafting needs a configured text provider; file staging, structured field extraction, and SQLite publication do not.
| Optional integration | Setup |
|---|---|
| LangSearch | Set LANGSEARCH_API_KEY for RFP web discovery. File imports and public directory sources remain available separately. |
| Puter | Install Node.js 24+, run npm --prefix scripts/puter ci --ignore-scripts, then set PUTER_AUTH_TOKEN. New course publications can generate cover art. |
| Gemini and Chroma | See provider setup and the Chroma security notes. Use the separate vector requirements only when needed. |
| OCR | Install Tesseract with English and Arabic language packs for scanned PDFs. Text PDFs, DOCX and Markdown do not require OCR. |
| Demo mode | Configure an external sample pack via REX_DEMO_EXAMPLES_DIR, then run python scripts/run_demo.py. The demo uses separate storage and ports 8010/8510. |
Restart FastAPI after configuration changes. See the provider guide for details.
.\.venv\Scripts\python.exe -m compileall -q backend frontend
.\.venv\Scripts\python.exe -m pip check
.\.venv\Scripts\python.exe -m pip_audit --local
.\.venv\Scripts\python.exe -m pytest tests/ backend/app/agents/rfp/tests/ -qThe latest application release check on commit 2d01734 recorded 514 passed, 7 skipped. Offline tests cover authorization, rejection and publication, retrieval and source withdrawal, orchestrator confirmation, RFP filtering, course covers, exports, and Streamlit interactions. Provider tests are opt-in; tests use isolated data and mocks where appropriate.
GitHub Actions runs syntax checks, a dependency audit, and the offline suite on Ubuntu with Python 3.14. CI does not deploy the application or verify provider availability.
| Path | Responsibility |
|---|---|
| frontend/ | Streamlit learner and staff workspace |
| backend/app/catalog/ | Uploads, staff review, publication, covers, and approved retrieval |
| backend/app/agents/ | Curriculum and RFP agent services |
| backend/app/instructor/ | Persistent chat, action planning, tool registry and confirmation |
| backend/app/documents/ · marketing/ | Document rendering and campaign drafts, reviews, and exports |
| tests/ · docs/ | Automated checks, architecture, workflows, and API contracts |
Read the current architecture, agent workflows, OpenAPI snapshot, and security and deployment notes for more detail. Team workstreams are documented in the WBS; individual contributions are traceable through Git history and pull requests.
This MVP is designed for controlled local use. Shared staff authentication is not individual identity or role-based access. A public deployment would need individual accounts, deployment security, monitoring and an operational data policy. Staff must verify real catalog data, procurement leads and accreditation criteria before using them in academy operations.