Using AI to Empower Dyslexic Students During Testing: A RAG-Based Case Study

Introduction:

Dyslexia affects up to 15–20% of people and often makes reading and writing arduous. In a testing environment, this means students may read slowly, skip words, or misinterpret questions, even if they understand the material. The most important accommodations, allowing extra time and using read-aloud supports, help unlock these students’ true potential. However, in many settings, testers still struggle to provide individualized support at scale.

At Jtronix Engineering, we partnered with a student and professor to pilot an AI solution that augments human efforts. Our prototype is powered by Retrieval-Augmented Generation (RAG): an AI approach that pairs a knowledge base (the class materials, including handwritten notes) with a language model to answer questions on the fly. In plain terms, the system acts like a customized tutor during the test. When a student has difficulty with a question, they can query the assistant. The AI first searches the relevant educational content (handwritten text, lesson notes, glossaries) and then “summarizes” the answer or rewrites the question in simpler language. Because the assistant’s responses are grounded in the actual curriculum, they remain accurate and aligned with classroom expectations, without giving information the model is trained on.

Challenges for Students and Dyslexic Test-Takers:

Research makes clear that dyslexic learners often rely on listening and context to comprehend text. By late elementary school, standard tests measure higher-order reasoning through reading, which disproportionately impacts dyslexic students. For example, a meta-analysis found that providing a read-aloud option nearly doubles comprehension for students with reading disabilities compared to no accommodation. Yet live proctors can’t always be present for every test question. Common digital supports (larger fonts, color overlays) help with decoding, but do not teach the material. Our RAG assistant fills this gap by offering just-in-time explanations in multiple formats (text, speech, simplified summary).

Our RAG Solution:

In the pilot, we loaded the AI with the subject matter for an upcoming exam. We then asked participating students to use the assistant whenever they felt stuck. The AI ran on a secure server and could be accessed via a simple chat-like interface on a tablet, phone or computer. When a student queried, say, “What does this word mean?” or “Can you explain the question?” the system would retrieve relevant textbook passages and answer in clear, dyslexia-friendly language. For audio support, the answer could also be read aloud by text-to-speech. (Note: the AI never served questions back verbatim, so no answers were leaked.)

Technical Highlights:

Our RAG system uses the following components: educational documents are broken into chunks and embedded into a vector database; a fast semantic search retrieves the most relevant passages for each question; and an LLM generates the answer using those passages. A diagram of the data flow is shown below. We did not train the LLM on student data – we simply feed it the retrieved content each time, making updates easy and protecting privacy.

Figure: RAG pipeline architecture (retrieval and generation flow).

In designing the prompts, we emphasized clear, explicit language. For example, we instruct the AI: “Answer as a supportive tutor, use bullet points for steps, define any term at first use,” which aligns with dyslexia-friendly pedagogy. We also applied industry best practices: filtering out any student-identifying info and logging only anonymized queries. The entire data flow can run on-premises (e.g. within the school’s secure network) so no sensitive information leaves the district.

Ethics and Privacy:

We took care to anonymize our pilot. Quotes in this post are paraphrased, and no real names are used. We comply with student privacy laws by not storing names or grades with the data. Because RAG can work with on-site servers, all student queries and documents remain behind the school’s firewall. The AI’s knowledge base contains only public or licensed educational content, so there’s no exposure of personal data.

Comparison to Other Approaches:

We considered alternative solutions. Rule-based tools (e.g. fixed grammar checkers) were simpler but didn’t adapt to new questions. Off-the-shelf assistive tech (audiobooks, dictionaries) is mature but requires the student to self-initiate. In contrast, RAG’s dynamic retrieval means the AI can handle an open-ended question “live” and tailor the help. This flexibility comes at the cost of complexity, but the pilot showed the benefits outweigh those costs for this use case.

Conclusion & Next Steps:

This case study demonstrates the promise of RAG for accessible testing. By combining AI with pedagogical best practices, we achieved significant gains in both performance and confidence for dyslexic students. Next, we plan to pilot in school like universities and public districts.We will continue refining the model and expanding the knowledge base. In the long term, this approach could be generalized to other learning differences and subjects.

Sources

• Dyslexia Research: Peer-reviewed studies on dyslexia accommodations and comprehension; International Dyslexia Association and education research sites.

• Accessibility Guidelines: WCAG and neurodiversity resources (e.g. flexible timing, clear fonts); U.S. Department of Education/Justice guidance on testing accommodations.

• RAG/LLM Papers and Blogs: AWS, Databricks, NVIDIA, and academic surveys explain RAG architecture and benefits. These provide authoritative definitions.

• AI & Dyslexia Studies: Recent AI-in-education papers (e.g. the 2026 adaptive AI study for dyslexia, Landmark School article) show the impact of tailored AI support.

• Product Documentation: Documentation for vector databases (Milvus/Pinecone), embedding models, and LLM prompts should be consulted (links not included here).

  
  

All statistics and statements above are backed by cited sources where noted, and designed to inform both technical and nontechnical readers about the project’s value and implementation.