훈맹정음

Hun-maeng-jeong-eum is a web service for university disability-support staff producing text-based learning materials for students with visual impairments. I built it independently and refined it with feedback from people around me. The service is currently being used.

The challenge was not simply obtaining an AI-generated transcription. Staff also needed to preserve the meaning of the source, check numbers and structure, correct errors, and deliver materials in the format each student could use. I brought these steps into one production-and-review workflow.

Feedback showed that preferences differ subtly between students, and that the same student can need a different structure in another course. I made these differences explicit in Markdown profiles instead of treating one formatting manual as a universal rule.

A profile defines page markers, headings and paragraphs, symbols, descriptions of tables and figures, mathematical notation, and filenames. Student and course tags locate the profile; the Markdown body contains the production rules. The planned selection flow matches both tags and falls back to a default profile when there is no match.

I connected AI agent APIs for production and review and introduced parallel processing for multiple lecture materials. The production agent writes the material using the source and the selected profile. Review is a separate role rather than the production agent simply endorsing its own output.

The specification gives the reviewer the source, profile, and current text, without the producer's identity or reasoning. Staff trigger a full review when needed. The review report identifies the location, source evidence, reason for the issue, and a proposed correction, including omissions, recognition errors, numbers, units, order, and profile compliance.

The planned workflow starts with a source file and a student–course profile. Production leads to a side-by-side source comparison, review, correction, and final TXT export. Staff can edit the text directly or request a targeted AI revision; a proposed revised version is previewed before being applied.

The specification also separates changes to a profile from changes to the current text. Existing work retains its applied profile version. Updating the profile changes subsequent revision and review rules without silently rewriting the material already produced. Saving and undoing are explicit actions.

The initial context is an internal university workflow. The specification supports reusable profile libraries so another institution can add its own production guidelines while keeping the common production-and-review process. Cross-institution sharing and shared workspaces remain expansion decisions.

The plan covers documents, slides, scans and images, PDFs, and audio transcription. These are specification targets rather than a claim that every file type and every planned screen is already available. The service's current use is confirmed; the detailed workflow described here follows the supplied project plan.