⚡ AAF Home Lecturer AI Property Dashboard ✉ Messages

Marking Lecturer Agent

Loading... Memory:
Step 1 Rubric
Step 2 Submissions
Step 3 Options
Review & Correct Grading Decisions

Mark each criterion correct (thumbs up) or wrong (thumbs down). Corrections train the RLHF engine for future runs.

Run grading first, then review decisions here.

Grading History

Click any run to view full results and explainability.

No history yet.

Learning Memory & RLHF

Loading...

RLHF Correction Patterns

Patterns injected into future grading prompts to correct model bias.

No patterns yet.

Student Chat Ask About Your Results

Students can ask questions about their own results. No rubric details or other students data are shared.

How to Use
  1. Upload your rubric (paste or upload .txt/.pdf/.docx/.pptx) in the Grade tab.
  2. Upload student submissions filename becomes the student ID. Supports 20+ formats.
  3. Set grading options and optionally enable plagiarism detection.
  4. Click Run Grading Pipeline. On CPU this takes several minutes per student.
  5. Review results each criterion shows score, evidence, and confidence level.
  6. Use thumbs up/down on each criterion. For thumbs down, enter correct marks and explain why.
  7. Click Submit Reviews to Memory this trains the RLHF engine for future runs.
  8. Use Student Chat tab to let students query their own results securely.
Frequently Asked Questions
Supported File Formats

Submissions and rubrics:

PDFDOCX/DOCTXT IPYNBPPTX/PPTZIP (multi-file) PNG/JPG (OCR)LaTeX (.tex) R Markdown (.rmd)MATLAB (.m) PythonR ScriptSQL HTMLMarkdownCSVXLSX
System Architecture

5-Agent Grading Pipeline

  1. Rubric Interpreter converts rubric into structured criteria with indicators.
  2. Submission Evaluator grades each criterion independently with document-aware section matching.
  3. Cohort Analysis computes statistics and applies moderate scaling if needed.
  4. Penalty Adjustment applies late, plagiarism, and collusion deductions.
  5. Feedback Generator writes structured academic feedback from the evidence collected.

RLHF Self-Improving Grading

Every thumbs-down correction creates a preference pair. When 3+ pairs exist for a criterion, a correction hint is injected into the LLM prompt: "This model tends to undermark by 2.1 marks adjust accordingly."

Local vs Cloud

Set LLM_MODE=local in .env for fully offline operation. Set LLM_MODE=online for OpenAI. All features work in both modes.