Correctable

Overview

An AI essay correction service deployed in high school English classrooms. Built as co-founder and CTO of EnglishAI, with ownership of architecture, the correction engine, and rollout into real schools.

Core Problem: A Perfect Correction Teaches Nothing

Asked to fix a student’s essay, a model returns flawless English — the worst possible output. The student sees writing that is not their own, cannot tell which changes mattered, and gains nothing transferable to the next essay.

Useful correction does the opposite: it leaves most of the writing untouched, targets only the errors blocking the student at their level, and explains the rule rather than merely applying it. Eliciting that behavior from a model is a prompt-and-evaluation problem rather than a capability problem — where most of the engineering went.

Key Contributions

• Led the full service architecture and development — from correction engine to classroom-facing application
• Built the GPT-4o based essay analysis and correction engine — structured so feedback attributes to specific spans rather than a rewritten whole
• Engineered pedagogical feedback prompts — constraining the model to correct selectively and explain, rather than rewrite to perfection
• Deployed and operated the service in high school settings — where the binding constraints are class time, teacher workload, and network reality, not benchmark scores

Tech Stack

• AI: OpenAI GPT-4o
• Frontend: React, TypeScript, Vite
• Backend: NestJS, TypeScript
• UI: AntDesign
• Infrastructure: AWS, GCP