Launch HN: Bloomy (YC S26) – AI-powered mastery learning for K-12

Alex Southmayd spent his days as a Teach For America 7th-grade English teacher trying to serve 30 kids with 30 different needs, and he never stopped thinking about Benjamin Bloom's old finding that one-on-one tutoring can push students roughly two standard deviations ahead of regular classroom instruction. That's the famous "2-sigma problem" — attention works, but it's never been affordable at scale. Southmayd's new company, Bloomy, is betting that cheaper AI models finally close that cost gap.

The product isn't a chatbot waiting for homework questions. Bloomy diagnoses a student's skill gaps, builds a personalized path through Math, English Language Arts, and Writing, and moves kids through three stages per skill: Base Camp for worked examples, Climb for guided practice with an AI tutor called BloomyBot, and Summit, a ten-question assessment with zero AI help where a student needs 90% to advance. Miss that bar too often, and the system reroutes you to a different skill entirely.

What's notable is how deliberately Bloomy restrains its own AI. BloomyBot follows a scripted tutoring ladder — ask what the student tried, nudge toward the concept, suggest a strategy, work through a step together, and only then offer heavier help. It's locked to the current lesson, can't operate during assessments, and doesn't get to decide whether a student has actually mastered something; that judgment stays with the deterministic mastery test, not the language model. Southmayd's reasoning is blunt: a model trying to be maximally helpful can be terrible at tutoring, since handing over the answer finishes the worksheet but skips the learning.

Early results are thin but suggestive. A pilot at a Massachusetts charter school with about 150 middle-schoolers showed roughly 1.8 times the expected NWEA MAP growth between winter and spring — an observational number, not a controlled trial, and Southmayd is careful to frame it that way rather than oversell it. Bloomy currently runs on a mix of Anthropic and OpenAI models, has zero-data-retention agreements with both, and says it won't let providers train on identifiable student data.

Pricing is straightforward: $39 a month for ELA, $19 for Writing Studio, with Math arriving July 31 at ELA's price point. Schools pay per student. Southmayd's path here — teaching, GMAT curriculum design, Lyft, a Stanford MBA, then AI work at McKinsey — reads like a founder who's been circling this exact problem for a decade before the tools caught up to the idea.