Client /
01EdTech AI
MagicSchool AI
We built the core systems that let MagicSchool scale to millions of teachers — every student interaction safe, fast, and affordable to run at global scale.
Read the caseMeasured resultValued at ~$500 millionThe number that mattered most.
How it works
Safe for every student. Fast for every teacher.
One system that keeps kids safe, keeps teachers moving, and keeps AI costs under control — at 75 million messages a month.
Safety pipeline / liveBuilt fordelivery map / 01
delivery map / 0101
Students chat with AI75 million messages a month
02
Real-time moderationMessage checked in under 10 seconds
03
Accuracy firstMultiple layers of checks resulting in < 5% false positives/negatives
04
Smart caching$30k saved monthly
346%more accurate moderation
$30k+saved monthly
<10 secper safety check
01 / Context
The challenge
MagicSchool had explosive growth. To win school contracts and keep it, all 75 million messages a month had to be safety-checked instantly — without slowing down classrooms or letting AI bills spiral.
02 / Delivery
What we did
- 01Shipped live safety in six weeks. Every message runs through layered checks in under 10 seconds — holding steady at 75 million messages a month without slowing classrooms.
- 02Made moderation 346% more accurate. Multiple layers cross-check each result, holding errors under 5% false positives/negatives, so schools can trust every decision.
- 03Cut AI bills by $30k+ a month with smart caching. Repeat checks are answered instantly from cache instead of re-paying the model every time, so costs stay flat as usage grows.
- 04Rebuilt voice for the classroom — faster speech-to-text and text-to-speech plus in-browser voice detection. 86% faster, even on weak school wifi, so lessons stay instant and don't cut out.
- 05Built document imports teachers actually use. Drop in 20 pages at once at 98% accuracy — 50 million+ imports so far — so old lesson plans move over in minutes, not hours.
- 06Shipped 1-click Drive and OneDrive imports. Teachers connect in seconds and pull lessons straight in — adopted by 74%, sharing in about half the time.
- 07Added proper math editing for STEM. Teachers can write equations correctly right in the lesson, so science and math teams don't need workarounds or leave the platform.
04 / Built with
Next.jsNode.jsWebAssemblyLLM pipelinesSpeech-to-text / Text-to-speechVercelLaTeX / KaTeXGoogle Drive / OneDriveNext.jsNode.jsWebAssemblyLLM pipelinesSpeech-to-text / Text-to-speechVercelLaTeX / KaTeXGoogle Drive / OneDriveNext.jsNode.jsWebAssemblyLLM pipelinesSpeech-to-text / Text-to-speechVercelLaTeX / KaTeXGoogle Drive / OneDrive
Ready when you are
Have something to ship?
New build or painful fix — send a few lines. You'll get an honest plan and cost range within 1 business day. No pitch, no pressure.