Problem
Most web platforms fail accessibility checks, and educational sites are no exception. People with dyslexia, ADHD or autism spectrum disorder (ASD) hit barriers that screen-reader fixes do not cover: dense text, low contrast, empty links and buttons. Single-LLM "WCAG repair" only patches alt text and contrast, and an LLM that generates a whole page differs on every run.
Example from the article: a user writes "I have ADHD and dyslexia, it is hard to read long texts" and uploads a WCAG fragment. The system sets the OpenDyslexic font, line spacing 2.0 and paragraphs of at most 3 lines.
Solution in brief
EMPI splits the work between three agents. One scores the complexity of the source text, one turns the user's free-text description into a profile with preset interface settings, and one fills a pre-written accessible HTML template with the adapted content. The LLM gets no freedom over the layout, which keeps output stable and WCAG-compliant.
Questions this case answers
- How to adapt the presentation of web content to a user's self-described cognitive profile with an LLM while keeping the output accessible and stable from run to run
- How to split an LLM personalization pipeline into separate agents for text analysis, user profiling and constrained generation
- How to adapt web page content to a user's cognitive profile (ADHD, dyslexia, ASD) with an LLM
- How to make LLM-generated interfaces accessible and stable from run to run
Tools and methods
- EMPI Model Protocol — JSON message exchange between agents, inspired by MCP's JSON-RPC
- textstat readability metrics — Flesch-Kincaid, Coleman-Liau, Dale-Chall, SMOG
- Profile presets — conditions mapped to fonts, spacing and chunking, based on DSM-5-TR
- Validated HTML templates — plus a JS validator for 8 WCAG 2.2 criteria
- LLMs — DeepSeek-V3 for synthetic profiles; five models in the evaluation
Google or a one-time email code. No password.