Backlogs form quickly
The workflow had to absorb sharp bursts rather than predictable daily volume.
Science Centre Singapore · Automation · Human-in-the-loop AI
AI-assisted assessment for school-holiday submission peaks.
A privacy-conscious proof of concept that reduces repetitive first-pass grading work while keeping final judgement, edits and Moodle updates under educator control.
Science Centre Singapore
01
Challenge
The issue was a concentrated holiday-period influx, not steady daily demand.
During school-holiday programmes, submissions arrive across courses in a short window while educator capacity remains fixed. Each record still requires opening the response, locating any submitted image, applying the rubric, drafting feedback and returning the result to Moodle.
First-pass effort repeats
Similar grading actions consume attention that is more valuable on exceptions and judgement.
Approved tools only
No Python, Power Apps or custom backend. The POC had to fit PAD, local files, the approved app and Moodle imports.
How might we absorb a temporary assessment backlog without removing educator judgement or expanding the data sent to AI?
02
Current journey
Without assistance, every submission repeats the same manual loop.
1Export
Pull worksheet data and submission artefacts.
2Match
Locate the correct response and image.
3Assess
Compare evidence with the rubric.
4Write
Prepare grade and feedback.
5Record
Return approved results to Moodle.
Moodle worksheet → local preparation → review → Moodle upload
03
Users
Three roles shape the service.
Workflow-based proto-personas derived from the proof-of-concept context, not formal research personas.
Science educator
Needs to clear a concentrated queue without giving up judgement or feedback quality.
Needs
- Fast review
- Consistent rubric use
- Final control
Friction
- Repetition
- Fatigue
- Manual transfer
Student participant
Needs timely, understandable feedback while the learning experience is still recent.
Needs
- Quick turnaround
- Clear rationale
- Useful comments
Friction
- Delayed results
- Inconsistent depth
Programme / LMS operations
Needs a controlled process compatible with approved systems and governance.
Needs
- Privacy
- Traceability
- Operational fit
Friction
- Restricted tools
- Fragile file exchange
04
Design principles
Automation supports judgement; it does not bypass it.
Human approval is mandatory
AI outputs remain recommendations until an educator reviews and saves them.
Minimise the data boundary
Only assessment evidence and grading instructions enter AI processing.
Fit the existing ecosystem
Use the LMS export/import formats rather than requiring a new platform.
Make changes traceable
Record prompt changes and human edits so the workflow can be reviewed and improved.
05
Service design
A staged workflow separates identity, AI processing and final grade handling.
Moodle & local
Export submissions
→
Match files to rows
→
Prepare CSV + image archive
AI boundary
Answer + image + rubric
→
Controlled batch API call
→
Suggested grade + feedback
Human controlled
Review / edit
→
Track changes
→
Upload approved grades
Sent for assessment
- Submitted answer text
- Optional submitted image
- Question / prompt
- Rubric / grading instructions
- Controlled output schema
Kept outside the AI payload
- Student name
- Student ID or email
- Course enrolment metadata
- Institutional grade record
- Unrelated personal data
The workflow in practice
The POC is a chain of controlled hand-offs, not one “autograder”.
The POC is a four-phase sequence — prepare, grade, review, return — that moves educator effort from repetitive first-pass production to verification, exceptions and final accountability.
Prepare · Phase 1
Validate CSV rows and submitted images before grading
Power Automate Desktop coordinates local preparation. The technically brittle part is preserving row integrity and matching each submitted image to the correct response before anything enters the AI-processing step. One grading record must resolve to the correct answer and image before any API request is created.
Grade · Phase 2
Load only the content needed for grading
The locally run application receives the response CSV and a zipped set of submission images, then exposes the grading plan before anything runs. Only the answer text, optional image, rubric and prompt enter the payload — names, IDs, email, enrolment and LMS metadata are excluded.
Grade · Phase 2
Make batch processing observable
The operator can check the model, limits and credentials, choose batch size and image handling, start the run and monitor progress and exceptions — rather than treating automation as a black box.
Grade · Phase 2
The AI output is a draft, not a record
Each result is a structured proposal: a suggested grade, a rationale tied to the response and rubric, editable student-facing feedback, and a local record key used to return the reviewed result to the correct LMS row.
Review · Phase 3
Keep the evidence and recommendation together
The reviewer sees the written response and associated image in one context. Grade and feedback remain editable so the educator can challenge the model rather than simply accept it.
Review · Phase 3
Human changes remain part of the record
Prompt configuration, proposed results and reviewer edits are treated as traceable workflow events. The approved result is then reformatted into Moodle’s expected offline-grading worksheet structure.
Return · Phase 4
Return only reviewed grades to the LMS
The approved result is reformatted into Moodle's expected offline-grading worksheet, selected for upload and confirmed explicitly. No unreviewed AI output is uploaded and the overwrite action stays visible, so the educator ends in the familiar LMS with clear confirmation that the human-reviewed grades have been applied.
06
Outcome & learning
The POC proved the workflow was feasible inside the constraints.
It demonstrated a way to reduce repetitive first-pass work, keep text and image context together, preserve educator ownership and limit the information shared with the AI step.
Operational value
Batch preparation and draft recommendations can reduce repeated grading actions during peak periods.
Trust and governance
Review, editing and change-log controls make AI use visible and accountable.
Technical feasibility
Approved tools and CSV interoperability were sufficient for an end-to-end round trip.
What was not measured
This was a proof of concept, not a production rollout. Time savings, grading agreement, educator satisfaction and peak-period throughput still require structured validation.
07
Product direction
Validate the value during the next holiday peak.
A limited pilot should test whether the workflow reduces backlog without weakening assessment quality or educator trust.
1Baseline
Measure manual time per submission and current turnaround.
2Pilot
Run one course or programme through the assisted workflow.
3Compare
Track edit rate, grade agreement, feedback quality and exceptions.
4Improve
Refine prompts, review navigation and error handling.
The intended outcome is not autonomous grading. It is a more resilient holiday-period assessment service where educators spend less time producing repetitive first drafts and more time applying judgement.
Reflection
The hard part was not the API call. It was designing a trustworthy operational system around it.
The project strengthened my approach to business analysis and product design: understand the peak-load problem, respect the real constraints, minimise sensitive data and make human control visible throughout the service.