Multi-agent smart academy
Five workflows. One reviewed source of course knowledge.
A collaborative platform for Andalusia Academy, connecting course advice, opportunity research, curriculum design, educational documents and marketing through persistent workflows and explicit review.
View sourceThe problem
Course facts can spread across uploaded documents, learner conversations, curricula and campaigns. Without publication and review boundaries, an unverified schedule or withdrawn course can continue to shape downstream content. The platform needs to coordinate useful AI workflows while keeping staff responsible for academy facts.
A reviewed course catalog anchors the system. The assistant retrieves published facts, generated artifacts retain their source snapshots, and domain services recheck approval and revocation conditions. A staff instructor interface can propose allowlisted actions, with confirmation for writes and separate review decisions.
Built with Faris Ahmed and Ahmed AlAbiad for Andalusia Academy. This case study follows main and its current README, architecture and team work breakdown.
Individual ownership, shared integration
English/Arabic advice, approved-only retrieval, source citations, guided learner qualification and persistent context. Deterministic level, availability, budget and delivery checks constrain recommendations. Lead capture requires consent.
Course document ingestion, native and optional scanned-document processing, extraction review, rejection/correction and explicit publication. Draft knowledge remains separate from published facts, with source withdrawal and revocation checks.
Ten template categories, English/Arabic DOCX/PDF, source-tracked generation and immutable revisions. Human review, approval/rejection and authenticated file-integrity checks connect outputs to approved course and curriculum data.
Assistant, course catalog/RAG and Document Generation
RFP Research and Curriculum
Marketing
System architecture, integration and testing are collaborative work. This role breakdown is documented in the main branch’s work breakdown structure and supplied team attribution.
Explore the system
Select a component to inspect its contract.
Catalog and RAG are also my primary workstreams. Architecture, integration and testing are collaborative. These nodes summarize responsibilities, with handoffs enforced by domain services.
One course through the platform
A source-aware walkthrough of the connected workflow.
Step 1 of 6
Staff provide a PDF, Word or Markdown course source. Optional Tesseract supports scans. Extraction produces candidates and warnings for inspection.
Engineering challenges
Extraction proposes course fields. Staff inspect warnings, verify facts, correct or reject sources, and explicitly publish an approved course. Draft or rejected material cannot silently enter learner retrieval.
Generated records retain course and curriculum snapshots. Current publication status, source hashes and review timestamps are checked again. A withdrawn or changed source invalidates downstream approval/export eligibility.
Instructor proposals and execution outcomes persist. Repeated executions return saved results, and a failed action stops later steps while preserving completed work. Cross-service activity remains outside a single atomic transaction.
Implementation & trade-offs
Optional Puter generation runs once per course revision after publication. Course browsing reads the stored image, and publication can succeed without cover generation.
Trade-offThis keeps catalog availability independent of an optional provider. Configuration and provider availability still need operator verification.
SQLite FTS5 retrieves published course facts without an embedding service or model download. Groq supplies the default text provider, with Gemini optional.
Trade-offThe default is inexpensive to run locally and easy to inspect. FTS5 is lexical search, so semantic matching requires the optional Gemini-embedding/Chroma path and its additional operational considerations.
The planner selects from 44 registered internal API tools. ASGI execution invokes existing authenticated routes and validators. Writes require confirmation, and review/publication decisions require a separate inspected action.
Trade-offThis reuses domain safeguards and limits execution authority. The planner’s implementation uses internal APIs. FastMCP exposes a separate ten-tool interface for external clients.
Ten template categories support source-tracked English/Arabic DOCX and PDF. Staff approve template versions, review drafts, request revisions and download authenticated files with integrity checks.
Trade-offEvery change becomes a new version, which improves traceability and preserves review history. Staff still need to inspect the content and sources before approval.
Persistent staff interactions, record review and internal tools share the application’s current staff authentication and declared reviewer identity.
Trade-offThe current model supports controlled local use. Individual staff accounts/RBAC, public-hosting controls and academy acceptance remain separate completion criteria.
Results & evaluation
Historical repository results.
Methods and limitations travel with the numbers.
The main README records 514 passed and 7 skipped at commit 2d01734. Offline checks cover authorization, source rejection and publication, retrieval and withdrawal, planner confirmation, RFP filters, course covers, exports and Streamlit interactions. Tests use isolated data and mocks; provider tests are opt-in.
Passing these checks supports the implemented paths and safeguards. Real academy acceptance, live-provider behavior and public-hosting controls remain separate work.
Limits & lessons
Connecting agents is also a data-lifecycle problem. Publication, source snapshots, confirmation and revocation need to be enforced in shared services so every interface follows the same review rules.
Inspect the work
Selected implementation and documentation reviewed for this case study.