Back to Systems

Case study // edumethod-ai

EduMethod AI

A learning workspace combining persistent learner context, pgvector retrieval, verification, and spaced repetition workflows.

Problem

A stateless assistant cannot preserve learner context, and ungrounded answers are difficult to audit against course material.

Constraints

  • Learner state and document context need consistent persistence.
  • Retrieval should remain close to the application data model.
  • Study workflows need deterministic scheduling rules.

Architecture

Syllabus Input -> Structured Content -> PostgreSQL/pgvector -> Learner Memory -> Verification -> Study Schedule

Data model

Learner profiles, source documents, embeddings, memory records, and review schedules are modeled as related application data.

Key decisions and trade-offs

  • Use pgvector alongside PostgreSQL to keep application and retrieval data within one database boundary.
  • Separate generated responses from verification state.
  • Use SM-2 scheduling logic for repeatable review intervals.

Failure modes

  • Missing source context should prevent a grounded answer from being implied.
  • Verification state remains distinct from the generated response.
  • Upload and retrieval steps can report failure independently.

Result

The project presents a data-centered approach to AI learning workflows rather than treating chat history as the product model.

Lesson learned

Persistence and verification are product capabilities, not optional metadata added after the model integration.