Skip to content

System card

How Q-MakerAI works — and what we measure

Q-MakerAI turns paper exams into structured, gradable practice. This page explains the architecture in plain terms, shows exactly which numbers we measure and how, and lists what we have not measured yet.

The pipeline

  1. 01

    Upload. A scanned exam PDF goes in. Long documents are decomposed client-side and streamed in chunks to serverless compute.

  2. 02

    Page analysis & region detection. Each page is classified, and question regions are detected as bounding boxes.

  3. 03

    Structured extraction. Each region becomes a structured item — passage, choices, answer key, formulas — in a versioned schema covering seven item types.

  4. 04

    Human review, twice. A region-review gate (drag to fix any boundary, source scan always in view) and a question-review gate (block-level editor) — nothing reaches a learner that a person has not approved.

  5. 05

    Variants & figures. Optional similar-question generation redraws each figure as executable plotting code rendered deterministically — not sampled from an image model — through 16 specialized figure engines, then screens every item through a four-axis quality gate (pHash, SSIM, CLIP, LLM judge) into PASS, REVIEW, or BLOCK.

  6. 06

    Deliver, grade, bank. One click deploys a test as a six-character code or QR. Grading applies computer-algebra equivalence (1/2 equals 0.5). Every item lands in Q-Bank for reuse.

Generative AI proposes; deterministic code renders; automated gates screen; humans approve. That composition — not any single model — is the product.

What we measure, and how

All figures below are measured on production infrastructure or on a defined internal test set (an 89-question golden set), with per-task billing logs. None are projections.

  • 98.9–100% per-stage pipeline success across 1,244 logged production calls (April–July 2026). This is technical reliability — not content accuracy.
  • ~10 minutes of AI processing turned an 8-page, 40-question paper into editable questions (measured, representative run).
  • ~35 seconds of AI processing per question on the 89-question golden set. Full-pipeline AI cost is about $0.083 per question, measured across production runs (April–July 2026).
  • Figure-render fallback measured at 4.8% — 3 of 63 figures kept the original crop instead of a redrawn one, on the CES demo set (5 subjects, July 2026). Figures the pipeline deliberately preserves are counted as intended behaviour, not failures.
  • ~85% lower cost per question to digitize an exam — roughly 6–7× cheaper than a documented 2024 outside authoring quote, based on our list pricing (a price-to-price calculation, not a measured customer outcome). The comparison covers digitization only; similar-question generation lands in the same band and is not quoted separately.

What we have not measured yet

  • Content-level extraction accuracy against an independent answer key. Our golden set currently verifies structure and parsing (89/89 items parse cleanly); field-level accuracy benchmarking is in progress, and we do not quote an accuracy percentage until it exists.
  • A representative QA-gate pass rate for figure regeneration. Early internal runs predate engine fixes and are not representative, so we do not cite them.
  • Learning-outcome efficacy. Q-MakerAI is an authoring and delivery tool; we make no claims about test-score improvement, and any such claim would require classroom studies we have not run.
  • Parts of the student viewer's accessibility roadmap are still in progress.

Student privacy

Students never create accounts. No passwords, emails, or phone numbers are collected from learners — a display name and a short session re-entry code are the only inputs. Educator accounts are email-verified, and every AI charge is itemized per task.

Intellectual property

One granted Korean patent (No. 10-2910853) and three pending applications cover the foundational technology — question-region detection, item-component tagging and structuring, and the production-and-review pipeline (applicant: DevApril Co., Ltd.).

Questions about our methods? The same review-first workflow shown here is what you'll see in the product.