EduCore is a multi-agent teaching and advising platform that runs entirely on institutional hardware — local model inference, local database, no cloud APIs and no third-party data processors. It was built for a context where FERPA and HIPAA compliance is not negotiable.
SACS enrolls working professionals across ten online graduate programs in two departments — Computer Science and Data Science, and Biomedical Data Science. These students attend live evening classes while holding full-time jobs. The result is a persistent advising gap: the questions arrive at 11pm, and the office hours are at 2pm.
The obvious fix — point a commercial chatbot at the course catalog — fails immediately on governance. Student records are protected under FERPA; a healthcare-adjacent institution adds HIPAA considerations. Sending student context to a third-party inference endpoint moves the compliance problem rather than solving it.
EduCore takes the other path. The model runs on institutional hardware. Retrieval happens against an institutional knowledge base. The platform is air-gap ready — fully operational offline once the model weights are downloaded — and there are no API keys to rotate, leak, or audit.
Design principle: the language model is treated as an untrusted component. Its inputs are filtered, its outputs are validated, and off-topic responses are independently detected — rather than trusting prompt instructions to hold. A human retains final authority at every decision point.
Four tiers, one machine, no egress. The browser talks to a FastAPI service, which orchestrates agent routing, retrieval, and generation against a locally hosted model and database.
Serving useful models on hardware a department can actually buy, and characterising the quality-per-gigabyte tradeoff honestly across model sizes.
Intent classification that decides which specialised agent handles a request, and branches to the right workflow rather than one prompt attempting everything.
Embedding and retrieving institutional knowledge so answers cite real course codes and real prerequisites, instead of hallucinating plausible ones.
Independent input filtering, output validation, PII/PHI detection and off-topic detection — engineered as separate layers, not as instructions inside a prompt.
Tool-use generation that guarantees well-formed JSON, so downstream exports to Blackboard, QTI and branded PDF are deterministic rather than best-effort.
Transparent hand-off from agent to human with category-based department routing, and a faculty dashboard that makes the queue visible and accountable.
Faculty and students need different things from the same knowledge base. EduCore separates them into task-specific agents with their own workflows, permissions, and safety posture.
This project takes motivated students at every academic level. The work is real: the platform is deployed, it has users, and the constraints are the ones you meet in production rather than the ones that appear in a course project.
Guardrail architecture and adversarial evaluation; agent orchestration design; retrieval quality and grounding metrics; systematic evaluation of locally hosted models against institutional tasks. Suited to students who want publishable systems work.
Data pipelines for catalog ingestion and knowledge-base maintenance; frontend interfaces for faculty and student workflows; export integrations with learning management systems; benchmark tooling and reproducibility scripts.
Foundations of how language models and retrieval actually work, guided evaluation of AI responses for accuracy and fairness, and structured exposure to how software gets designed, reviewed and shipped.
How to apply. Send a CV and a short statement of interest describing relevant experience and which of the six technical areas above draws you. Prior AI experience is welcome but not required — clear thinking and follow-through matter more.
Researchers looking for more detail on the architecture, the guardrail evaluation, or collaboration are welcome to write to the same addresses.
The platform scales down to hardware a single department can own. Model size is the main dial, and the tradeoff is stated plainly rather than hidden.
| Model | VRAM | Representative hardware | Quality |
|---|---|---|---|
qwen2.5:72b | ~42 GB | 1× A6000 or 2× RTX 4090 | Best |
qwen2.5:32b | ~20 GB | 1× RTX 4090 | Good |
qwen2.5:14b | ~10 GB | 1× RTX 3080 or better | Moderate |
qwen2.5:7b | ~5 GB | Any modern GPU | Basic |
| Layer | Technology |
|---|---|
| Frontend | React 18 · TypeScript · Vite · TailwindCSS |
| Backend | FastAPI · Python 3.11+ · Pydantic v2 · Beanie ODM |
| Database | MongoDB 7.0 |
| Inference | Ollama · Qwen 2.5 family |
| Authentication | JWT with Argon2 password hashing · role-based access |
| Deployment | Docker Compose · air-gap capable |
EduCore is developed in the DISCS Lab within the Department of Computer Science and Data Science at the School of Applied Computational Sciences, Meharry Medical College — a historically Black institution in Nashville, Tennessee.
Directs the research programme and sets the governance and privacy requirements the architecture is built to satisfy.
Leads the research direction across the agent architecture, guardrail design and evaluation methodology.
Built the platform end to end — architecture, backend, frontend, guardrail implementation and deployment.
Leads testing and validation of the platform, including adversarial evaluation of the guardrail layers.
The design constraint that shaped everything: institutional data stays institutional. EduCore exists because the compliance requirement came first and the architecture followed from it, rather than the reverse.