index c35039f..2afcbe0 100644
@@ -1,7 +1,13 @@
---
+first_captured: 2026-04-10
+sources:
+- sources/google-sheets-ideas.md
status: raw
tags:
- software
+- education
+- ml
+- writing
title: college essay grader
type: idea
updated: 2026-04-11
@@ -10,4 +16,8 @@ visibility: public
# college essay grader
-vector embedding distance for essay quality.
\ No newline at end of file
+a tool that grades college essays using vector embedding distance as a proxy for quality. the core technical insight: you can measure how close an essay's embedding is to high-quality exemplar essays (accepted essays, award-winning writing) versus low-quality exemplars, and use that distance as a quality signal. this is a more principled approach than rubric-based grading, because it captures holistic quality that is hard to reduce to discrete dimensions.
+
+the practical implementation would be: collect a labeled dataset of college essays (accepted/rejected, scored by counselors), embed them with a quality text model, and train a scoring model on top of those embeddings. the scores could be broken down by sub-dimensions — distinctiveness, voice, structure, relevance — each with their own embedding cluster. a key product decision: do you show a score or show "your essay sounds like these other essays" comparisons? the comparison view might be more actionable because it surfaces the fix, not just the diagnosis. connects to [[ultimate-describer|precision description engine]] for the making-language-more-precise angle and to [[writing-tools|writing tools suite]] for the broader writing quality product space.
+
+the market context: college counseling is expensive and access to good essay feedback is highly unequal. a cheap, good-enough automated grader could meaningfully democratize college prep. it also has a clear wedge into the [[cluster-learning-education|learning and education]] cluster — college prep is high-stakes and underserved by AI tools that are either too generic (ChatGPT) or too expensive (private counselors). connects to [[youre-not-behind|you're not behind machine]] for the college anxiety angle and to [[oncue|OnCue]] for the interview/application prep overlap.
\ No newline at end of file