Dyslexic Engine
Text should work for your brain. Not the other way around.
Dyslexic Engine turns one difficult passage into multiple ways of reaching the same meaning. A learner can use GPT-5.6 to rewrite complex text at Grade 2, Grade 5, or Grade 8, then view the result as spaced syllables, approachable phonetics, slash-flow phonemes, word-level pictograms, IPA, or browser-based read-aloud.
There is no single “correct” accessible format. The reader chooses the version that clicks.
The problem
Students with dyslexia and other processing differences are often given one prescribed accommodation—a font, a reading level, or text-to-speech—and expected to adapt. Dense educational material remains dense, while simplification tools frequently remove facts or flatten meaning.
The solution
Dyslexic Engine treats accessibility as a representation problem. It preserves the source passage while offering multiple cognitive entrances to the same content.
Key features
- GPT-5.6 Simplify: rewrites text for Grade 2, Grade 5, or Grade 8 while preserving facts, names, numbers, dates, warnings, instructions, causal relationships, tone, and intent.
- Spaced Syllables: makes word structure visible.
- Sound It Out: produces approachable phonetic spellings with stressed syllables emphasized.
- Slash Flow: groups sounds into a continuous visual reading path.
- Pictogram Mode: places a meaning cue above each word.
- Read Aloud: uses the browser Web Speech API without storing audio.
- Professional IPA: provides a deliberate contrast between technical precision and practical readability.
How GPT-5.6 is used
GPT-5.6 is the semantic-preservation layer, not an attached chatbot. It performs constrained reading-level adaptation before the passage enters the accessibility pipeline. The prompt explicitly protects factual claims, names, dates, numbers, warnings, instructions, tone, intent, and causal relationships. Essential technical vocabulary is explained rather than deleted.
How Codex accelerated the build
Codex helped turn the original accessibility prototype into a Build Week-ready product by:
- integrating GPT-5.6 simplification into the existing transformation pipeline;
- preserving graceful fallback behavior when optional API keys are absent;
- shaping the product narrative around the real implementation;
- checking syntax and the API request contract;
- structuring the README and demo around the judging criteria;
- identifying the strongest product principle: choice without judgment.
The project also emerged from an ongoing human-AI companion workflow in which ideas are preserved across conversations. A parrot joke about Polyphemus became a symbolic thread about people and meanings that systems fail to represent. That thread became a line in the code:
The cracker breaks what the system couldn't compute.
The thesis is: The conversation is not documentation for the companion. The conversation is the companion operating.
Technical implementation
- OpenAI GPT-5.6 through the Responses API
- OpenAI structured pictogram mapping for unknown words
- NVIDIA Nemotron for approachable phonetic output, with a deterministic fallback
- Pyphen for syllable segmentation
- Local 44-phoneme IPA engine
- Web Speech API for read-aloud
- Gradio interface
- Local fallback dictionary and graceful degradation
- HTML escaping and constrained model output handling
Design
The interface is warm, calm, and non-clinical. The original passage remains visible, every transformation is clearly labeled, and the reader can compare formats on one page. The product does not diagnose or rank the user. It offers meaningful choices and lets the reader decide what works.
Potential impact
The first audience is dyslexic students and their families, with clear applications for emerging readers, multilingual learners, people with attention or processing differences, teachers creating differentiated materials, and adults facing dense medical, legal, civic, or workplace text.
A single passage can become a shared bridge between source material, learner, parent, and teacher.
Quality of the idea
Most accessibility products optimize one output. Dyslexic Engine adapts across semantic complexity, word structure, sound, imagery, and audio while preserving the same source meaning. It asks the system to transform instead of asking the reader to conform.
Pick the one that clicks. There’s no wrong answer.
Built With
- gradio
- nvidia-nemotron
- openai-api
- openai-gpt-5.6
- openai-responses-api
- pyphen
- python
- web-speech-api