Mulyayan AI
Mulyayan AI is a teacher-governed visual assessment workspace for handwritten student work.
Teachers can submit a consent-safe handwritten image or page-based PDF, optionally with a marking scheme. The system preserves original work immutably, proposes evidence-linked semantic overlays, and renders a teacher-style red-ink underline, circle, arrow, and correction beside the relevant handwritten line.
The live demo focuses on algebra. In the golden case, a student expands 2(x + 3) as 2x + 3. Mulyayan AI links the proposed correction to a rubric criterion, visible evidence, normalized source geometry, confidence, and review state. A teacher can approve, reject, edit, reposition, approve the page, and only then release feedback to the learner. The learner receives a targeted repair activity, while the teacher sees approved-only formative patterns without ranking students.
Safety and privacy
Original page assets are immutable. Annotations are separate semantic overlays. Student views never expose suggested or rejected annotations, confidence values, raw analysis, or teacher edit history. Ambiguous, malformed, or low-confidence model output fails into teacher review rather than student-visible feedback.
How GPT-5.6 and Codex were used
Mulyayan AI includes a server-side GPT-5.6 structured-analysis provider. It is designed to send the minimum necessary page context and rubric, then request a strict RubricAnalysis object containing closed-taxonomy decisions, visible evidence, normalized anchors, confidence, and a bounded repair recommendation. Deterministic application policy validates schema, rubric IDs, geometry, confidence, and safety language before an annotation can render.
Codex was used extensively to build the typed provider boundary, FastAPI service layer, SQLite persistence and audit trail, immutable upload path, semantic SVG red-ink renderer, teacher-review canvas, student repair flow, evaluation harness, tests, documentation, and submission materials. Runtime never executes model-generated code.
Scope
The submitted end-to-end demonstration is algebra. The repository also includes a versioned literature subject-pack contract, but it is explicitly not presented as a validated live grading claim. All included fixtures are synthetic or consent-safe.
Built With
- codex
- fastapi
- gpt-5.6
- next.js
- openai-api
- pydantic
- python
- react
- sqlite
- svg
- typescript
