Education technology · Korea · 2020 to 2024 · KiwiOPIC · 2020 to 2024
A test-prep company that now generates its own curriculum: about 80% of the manual work handled by AI.
Hired to write English lessons. Left having turned the company's production line into an AI-assisted pipeline.
~80%of the manual curriculum work now handled by AIPer Alec, confirmed 2026-07-25; the company's own estimate.
95%+transcription and feedback accuracy on learners' spoken answers
Fig. · The same production line, before and after. Width is share of the work.
What was broken
Before
Every practice question, model answer, and lesson for a computer-delivered oral proficiency interview was written by hand by a small team. I was hired to write that curriculum.
What I built
The build
Question generation: a pipeline that produces OPIc-style interview questions across the exam's topics and difficulty levels, in the register the test actually uses.
Curriculum generation: lesson and test material drafted, formatted, and run through a review cycle, with the model taking the first pass and a human the last.
A speaking feedback loop: speech-to-text wired to a language model so learners' spoken answers were transcribed and scored, at 95%+ accuracy.
Process automation between drafting and publishing, and the team trained on the new workflows as they were built.
What changed
Before and after
Before
By hand
Writing questionsBy hand, one at a time
Production timeBaseline
FeedbackA teacher, days later
HeadcountWould have grown
After
The system
Writing questionsGenerated per topic and level, reviewed
Production timeAbout 40% less
FeedbackTranscribed and scored, 95%+
Headcount200+ curriculum hours shipped without hiring
Where the numbers come from. These figures are the company's own estimates as Alec reported them; there is no primary artifact in our records. Treat the percentages as approximate.
For engineers
How it's built
Architecture
Prompt pipelines per content type, a human-last review cycle, and a transcription-plus-scoring loop for spoken answers.
Where it runs
Inside the company's own tooling at the time.
Tests and evals
Model-first, human-last review on every published item; accuracy tracked on the feedback loop.
Cost and latency
Not recorded.
Guardrails
A human signed off on every published lesson.
The tradeoff we chose
Why this way and not the other
Generate the first draft, never the final. The register of the test is narrow enough that a model drafts it well and a teacher's last pass keeps it honest.