AccessLens

AccessLens - Learn & Grow Project Track

Inspiration

For the past four years, I have taught and worked closely with children with special needs. During that time, I have seen intelligent, curious students understand a concept during a conversation, then struggle when the same idea appeared inside a dense worksheet or complicated digital assignment.

They were not struggling because they lacked the ability to learn. They were spending their energy decoding long instructions, remembering information between screens, switching between multiple tools, working under time pressure, or responding in a format that did not match how they communicated best. When this happens repeatedly, students can begin to believe that they are the problem. A barrier in an assignment slowly becomes, “I’m not smart enough,” or “I’m just bad at school.”

I also saw how quickly that story could change when the material was presented differently. Breaking instructions into visible steps, keeping important information available, reducing unnecessary switching, or offering another way to respond could reveal understanding that had been there all along. The academic content did not become easier; the student was simply given a fairer path to demonstrate what they knew.

That experience inspired AccessLens. I wanted to help educators recognize the invisible cognitive and functional demands inside their assignments before students are forced to struggle through them. AccessLens is built around one principle: remove the barrier, not the challenge. Students should still think deeply and complete meaningful work, but no student should mistake an inaccessible assignment for a limit on their intelligence.

What it does

AccessLens is an extraneous cognitive and functional load manager for digital learning. It acts as an accessibility preflight for assignments, helping educators identify demands that consume a student’s limited attention and working memory without contributing to the intended learning.

Traditional accessibility checkers evaluate individual pages for issues such as contrast, headings, alt text, or keyboard support. Those checks are essential, but they do not reveal what happens across the complete student journey. Every individual page could technically pass while the assignment still requires a student to memorize results, switch between several tools, relocate instructions, work against a timer, drag small objects, or respond through a single modality. AccessLens analyzes the task itself.

An educator begins by pasting assignment instructions or uploading a text-based PDF. AccessLens identifies possible learning objectives, but the educator must confirm and lock the real objective before analysis begins. This creates a clear boundary between two kinds of difficulty:

Essential load: the intellectual work the learning objective is meant to assess.
Extraneous load: difficulty introduced by how the assignment is presented, organized, or completed.

For example, reasoning about how concentration affects molecular movement may be essential to a biology assignment. Remembering several simulation values while switching to another tab is not. Explaining a scientific relationship may be essential; dragging an answer into a small target or recording that explanation only through speech may be incidental.

Once the goal is locked, AccessLens reconstructs the assignment as a step-by-step student journey. For every step, it identifies:

The action the student performs
The environment in which it happens
Information produced or needed later
Whether that information remains visible
Working-memory demand
Context switching
Reading and instruction load
Fine-motor precision
Time pressure
Sensory dependencies
Required communication modality
Relevance to the learning objective

AccessLens then uses a deterministic engine to calculate measurable indicators of friction. Its working-memory analysis tracks information that must remain mentally “live” across multiple steps and compares the peak demand with research suggesting a capacity of roughly four chunks without rehearsal or external support, based on Cowan’s work.

Its context-switching analysis counts transitions between environments and detects repeated back-and-forth movement. This is informed by Monsell’s research on switch costs and Altmann and Trafton’s work showing how goals decay during interruptions and intervening tasks.

The timing analysis estimates reading and action time using Brysbaert’s adult nonfiction reading-rate research. It calculates both a mean reading estimate and a more conservative estimate, showing when a time limit may measure processing speed instead of subject knowledge.

AccessLens also examines instruction density, readability, forced response formats, sensory-only information, and precision-based interactions. These checks reflect research and principles including Sweller’s cognitive load theory, Fitts’s law, assessment validity, WCAG 2.2, W3C COGA guidance, and CAST’s Universal Design for Learning Guidelines.

The result is not presented as a black-box AI opinion or a WCAG compliance score. Educators receive a visual journey showing where friction occurs, which steps are affected, how severe the barrier may be, and whether the demand is essential, related, incidental, or still uncertain.

Every barrier includes a trace back to the exact language in the original assignment. AccessLens highlights the source passage, explains why it was flagged, connects it to the learning objective, and presents the relevant accessibility or research guidance.

Educators can then review a repair plan. They can apply, reject, or customize each recommendation. Every repair includes a rigor check explaining what academic evidence remains intact. AccessLens can also test the journey under functional access conditions such as reduced working-memory availability, keyboard-only navigation, limited motor precision, no spoken response, no audio, no color information, or additional processing-time needs.

Finally, educators see the original and revised assignments side by side. Only the repairs they approve are incorporated. They can review what changed, understand why it changed, and export the improved assignment. AccessLens does not diagnose students or replace educator judgment. It gives educators a structured, evidence-informed way to see demands that may otherwise remain invisible.

How we built it

I built AccessLens with Next.js 16, React 19, TypeScript, Tailwind CSS, and the Anthropic API. The technical architecture deliberately separates interpretation from measurement. Claude reads an assignment and converts it into a typed task graph. That graph describes the sequence of student actions, environments, information dependencies, functional demands, friction moments, and connections to the locked learning objective.

The model is not allowed to create scores, measurements, or citations. After the task graph is validated with Zod, a deterministic TypeScript engine calculates the results. The same task graph will always produce the same working-memory, switching, timing, readability, and friction measurements.

To make the AI analysis more trustworthy, I built an evidence guard that checks every quotation against the original assignment. If a quotation cannot be verified, AccessLens discards it instead of showing potentially fabricated evidence. Accessibility citations are also selected from a curated standards library rather than generated by the model.

For PDF support, PDF.js extracts the document’s text locally in the browser. I built logic to reconstruct positioned text fragments and rejoin sentences broken by PDF line wrapping. This preserves the original wording so evidence can still be verified and highlighted.

The interface guides educators through a deliberate workflow: add an assignment, lock the goal, scan the student journey, trace barriers, review repairs, test functional conditions, preview the student experience, and export the result.

Because AccessLens is itself an accessibility product, I also implemented keyboard navigation, focus management, screen-reader announcements, semantic page structure, visible interaction states, responsive layouts, and reduced-motion support.

Challenges I ran into

The hardest part was separating real academic challenge from difficulty caused by the assignment’s design. Locking the learning goal gave AccessLens a clear way to judge whether each demand was essential or unnecessary. We also needed the AI to be trustworthy. Claude interprets the assignment, but a separate engine calculates the measurements. Every highlighted quote is checked against the original text, so educators can verify each finding. PDFs were another challenge because they often store text as disconnected pieces. We also had to present a detailed analysis without overwhelming educators, so the workflow reveals information one step at a time.

Accomplishments that I'm proud of

I’m proud that AccessLens looks beyond basic accessibility checks and examines the full experience of completing an assignment. Every barrier can be traced to the original instructions, and every suggested repair is checked against the learning goal. Educators can also test different access needs and see which steps may cause problems. Most importantly, the educator stays in control. They can review, customize, or reject every change before exporting the revised assignment.

What I learned

I learned that accessibility is often a property of a sequence, not just a screen. A value that is visible in one tool but needed later in another creates a memory barrier between two otherwise accessible pages. A clear instruction can still become difficult when it requires repeated context switching. A valid question can become an invalid assessment when the required response format introduces an unrelated barrier.

The research also reinforced that reducing extraneous load is not the same as making learning easier. It means protecting a student’s cognitive resources for the work that matters. The goal is not to eliminate productive struggle; it is to prevent irrelevant demands from competing with learning.

Technically, I learned that reliable educational AI needs strong boundaries. Structured schemas, deterministic calculations, verified source evidence, curated citations, and human review were as important as the model itself.

What's next for AccessLens

The next step is to evaluate AccessLens with educators, special-education professionals, instructional designers, and disabled students. Their feedback will help validate whether the identified barriers reflect real classroom experiences and whether the suggested repairs are practical. I also want to add direct integrations with learning management systems so educators can analyze assignments before publishing them. Future versions could support additional document formats, reusable repair templates, organization-level accessibility patterns, and before-and-after comparisons that quantify how much extraneous load was removed. Longer term, AccessLens could become a design system for inclusive instruction rather than only an assignment analyzer. It could help educators build accessible tasks from the beginning, monitor an entire course journey, and recognize recurring barriers across lessons. The goal is to make accessibility preflight as normal in education as spell-checking: a routine step that helps every student spend less effort navigating the assignment and more effort learning.