01
Generate
Five presets + free text
- Basic · Graph · Applied
- Proof / essay
- Killer-question mode
- Natural-language requests
EXAMFORGE · KOREAN KSAT MATH AI
An AI that learned past papers. It writes new math questions.
Full KSAT math curriculum in one model. Presets, multi-LLM generation, 5-axis auto-verification — to PDF.
Full KSAT curriculumMulti-LLM routing5-axis auto-verify
The real problem in Korean math academies
Good math questions run thin every week.
Past papers are finite. Annual N-books drain fast. The 2025 cartel crackdown cut off the human-author supply line. And ChatGPT prompting is inconsistent on answers and solutions.
01
Limits of the past-paper bank
KICE and provincial mocks are a fixed pool. Every year the same thread appears on Orbi: "Done with every N-book — now what?"
02
Cartel crackdown cut the supply line
In Dec 2025, prosecutors indicted 249 teachers and top instructors over ₩21.3B in off-the-books question trades. That supply line is gone.
03
Mathflat curates — it does not generate
9,500 academies pay Mathflat ₩200–300K/month for question-bank curation. New questions still have to be written by hand.
04
ChatGPT prompting is inconsistent
Right answer? Rigorous solution? Correct difficulty? A human has to re-check every item. 30 minutes becomes an hour.
Why now
Three reasons this window is open.
2024 private-ed spend
₩29.2T
All-time high · per-student spend up 7.7% YoY
AIDT usage rate
8.1%
Public-school AI failed → demand shifts to academies
2028 KSAT
AI-authored items
KICE formalizes AI authoring · policy favors teacher-side AI tools
What ExamForge does
Preset → Generate → Verify. One set per week, done.
01
Preset
five built-in modes
Basic · Graph · Applied · Proof/essay · Killer. Pick a subject, unit, and difficulty; tap a preset and the style is set. Additional requests go in as natural language.
02
Generate
multi-LLM routing
OpenAI · Anthropic · Google · Groq — four providers in one panel. Structural immunity to single-vendor outages, pricing shifts, and policy changes.
03
Verify
5-axis automation
Every item must pass answer correctness, distractor quality, Korean fluency, solution rigor, and difficulty. Failed items never enter the bank.
Academic and product evidence
Data and a live pipeline — not adjectives.
Core screens
The actual product.
02
5-axis verify
Every item must pass
- Answer correctness
- Distractor quality
- Korean fluency
- Solution & difficulty
03
Question bank
Past papers + generated
- Subject · unit · level
- Source flag (past paper)
- Tags · full-text search
- LaTeX rendering
04
Sets · PDF
Exam-ready output
- Default exam format
- Bulk register
- Role-based quotas
- Copyright hygiene
Why not a ChatGPT wrapper
Three things a wrapper can't copy.
01
5-axis auto-verification
Answer, distractor quality, Korean fluency, solution rigor, and difficulty — checked on every item. Fail any axis and it never ships. ChatGPT writes "plausible" — we ship "verified."
02
Multi-LLM routing
OpenAI · Anthropic · Google · Groq in parallel. Structural immunity to pricing, quality, outages, and policy drift. Single-vendor wrappers can't match this.
03
KSAT curriculum tree
Math I · II · Calculus · Stats · Geometry modeled as a tree. Unit-level generation, bank filters, and difficulty distributions all operate at the unit layer.
Legal · safe
Copyright, policy, quality — all three.
Copyright-clean generation
Not redistribution — style learning and regeneration. Aligned with the 2025 MCST generative-AI guideline.
Cartel-safe by design
No human authors in the loop — structurally outside the "question trade" regulations.
Human approval before deployment
5-axis auto-verify plus instructor or editor sign-off. Unapproved items never reach students.
Pricing
Priced for the way you operate.
Mathflat benchmarks at ₩200–300K/month. ExamForge starts at the same anchor with generation and verification included.
Most picked
Standard
290,000₩/mo
Academy · 30–100 students
- 500 items/month
- 500MB storage
- 5-axis verification report
- PDF exam format
- Team-shared bank
Publisher
990,000₩/mo
Publishers · large academies
- Unlimited generation
- Unlimited storage
- Bulk register · bulk export
- Custom curriculum tree
- White-label option
Enterprise
Franchises · districts · publishing groups
Per-branch licensing, dedicated curriculum, on-prem deployment, SLA. Custom scoped.
Rollout
Four weeks, one academy transformed.
- 01
Consult — scope your subjects and weekly workflow (30 min)
- 02
Setup — register your curriculum tree and team accounts
- 03
Pilot — one instructor authors a full week's set with ExamForge
- 04
Launch — rollout to all classes with weekly reports
FAQ
Answers up front.
Are AI math items actually classroom-quality?
Every item must clear 5-axis auto-verification — answer, distractor, Korean fluency, solution, difficulty — before entering the bank. In a 2025 field study across 91 courses and 1,686 students (arXiv:2508.08314), AI-generated items beat expert items on IRT discrimination.
How is this different from Mathflat?
Mathflat curates a search-driven question bank. ExamForge generates new items by unit, difficulty, and preset. Many academies run both tools side by side.
Why multi-LLM?
Single-vendor lock-in passes pricing, outage, and policy risk straight to the academy. We route across OpenAI, Anthropic, Google, and Groq and compare outputs per item.
What about copyright?
We don't redistribute originals — we learn style and regenerate. Aligned with the 2025 MCST generative-AI guideline. Every item ships with a verification report you can keep.
Do you support non-math subjects?
The full KSAT math curriculum is modeled today. English, Korean, and science are rolling out; Publisher-plan customers can customize the curriculum tree directly.
Academies · publishers · top instructors
One verified set per week. The supply problem ends.
Thirty-minute call. We scope a pilot to your subjects and academy size.
Reply within 24 hours · 3–4 week launch