supe-log/writing-engine

Writing Engine is a tool that grades student essays the way a trained human rater would, rather than writing essays for students. You feed it a piece of writing scored against an official rubric, such as the STAAR state writing test used in schools, and it produces a score along with an explanation for that score backed by evidence from the text itself. It never generates essays on a student's behalf, only assessments and feedback. The project has three connected parts. The first is the scoring engine itself, which is checked against real, already scored essays from an actual state exam to measure how closely its grades match a human grader, including how confident that match is. The second is a builder process that can create new scoring engines for other rubrics and example essays, packaged as a reusable skill. The third is a background agent that runs the scoring engine on live incoming material, such as a stream of new essays or public weather alerts used as a stand in data feed, but that agent is deliberately blocked from writing or publishing anything until it has proven, with evidence, that it deserves that permission. A key idea running through the whole project is that each time it runs, it compares its own output against the fixed rubric, keeps only the changes that measurably improve its score, and applies those lessons the next time it runs, so its accuracy is meant to improve over repeated cycles rather than stay fixed. Every prompt and piece of content going in or out is also scanned for security issues and personal information before being processed. The project is written in TypeScript and requires Node.js 20 or newer with npm. A full demo and benchmark can be run entirely offline on prepared sample data with no API keys needed, and there are separate commands for running the demo, running the benchmark, running the test suite, and running a single live cycle. It can also be configured to run against different underlying AI models, including a self hosted option, by changing environment variables rather than code. The project was built for a hackathon.