AIdeas: CogniPath
Social Impact — educational accessibility for students with visual disabilities, dyslexia, and ADHD in public schools across Latin America.
AIdeas: CogniPath — The barrier was the format
My Vision
I started thinking about CogniPath when I noticed something that sounds obvious but has enormous consequences: most science educational materials are designed for one format. Plain text, static graphs, formulas that assume everyone can see them and process them the same way.
There are brilliant students in public schools in Argentina, Colombia, and Mexico who fall behind in STEM content not because they can't understand it, but because the format it arrives in doesn't work for their brain or their senses. A student with a visual disability faces an enzyme activity graph and there's nothing — no available system describes the curve with the level of detail they need to answer an exam question. A student with dyslexia reads the same paragraph four times and the text is still an impenetrable block. A student with ADHD has the knowledge, but the activation energy to start a study session is so high they never begin.
CogniPath is an educational format transducer. It takes any STEM material — graph, text, formula, diagram — and converts it to the format that specific student actually needs: structured audio description across 4 navigable levels, reformatted text with Bionic Reading, AI-generated pictograms for scientific vocabulary, a Socratic tutor that guides without giving direct answers, and a voice command system that lets you control the entire application without touching the screen.
What I built for this challenge is a fully functional prototype: a single-page application demonstrating each of these modules working in real time, using the Web Speech API for native browser voice synthesis and recognition, with an architecture designed to connect to Amazon Bedrock, Amazon Polly Neural, and Amazon Transcribe Streaming in production.
Why This Matters
There's a number that stuck with me when I was researching this project: a systematic review of 258 studies on AI in STEM education (arXiv, November 2025) found that adaptive content tools have an effect size of d=0.45–0.70 on comprehension and retention for neurodivergent learners. That's a large effect. And yet the same study identifies the most critical gap in the field as the lack of real integration between augmentative and alternative communication (AAC) systems and STEM educational platforms.
That gap has a concrete name. ARASAAC, the most scientifically validated pictogram system in the Spanish-speaking world, has 18,000 pictograms. But it doesn't have "covalent bond." It doesn't have "G2 phase of the cell cycle." It doesn't have "partial derivative." Scientific vocabulary at the secondary and university level simply doesn't exist in any standard pictographic system.
What CogniPath does is close that gap by combining ARASAAC with Amazon Titan Image Generator: for every STEM concept that doesn't exist in the validated corpus, it generates a pictogram on the spot, validates it with Claude 3.5 Sonnet to verify the representation is accurate, and caches it in DynamoDB. The second time anyone queries that concept, the cost is zero.
For students with visual disabilities, the Spatial-Audio Worker transforms any graph into a 4-level navigable description: first the general context ("line graph showing the relationship between temperature and enzyme activity"), then the axis structure, then the main trends, and finally the exact numerical data as a table. The student navigates with keyboard arrows or voice commands, decides how much detail they need, and can replay any level. It's the auditory equivalent of zooming into a graph.
For students with ADHD, the Narrated Pomodoro turns invisible time into audible time. The app doesn't just count down — it talks. At the ten-minute mark it says "We've been studying for ten minutes. You're doing great." When five minutes remain it says "Almost there." When the session ends it saves exactly where you stopped and tells you the concept you were working on. And during the break: total silence. That decision — active silence during the break — is grounded in research on cognitive recovery in ADHD. An app that keeps talking during the "break" doesn't produce real recovery.
The voice command system solves something subtler: poor discoverability, which according to AufaitUX (2025) research is the number one reason for early dropout in voice interfaces. Users don't know what they can say. In CogniPath, when you activate the microphone and say "help," the app narrates exactly what you can do from wherever you are. Every screen has its own contextual commands.
I work from Salta, in a province in northwest Argentina where the reality of public schools isn't abstract. I'm not solving a hypothetical problem.
How I Built This
The Agent Architecture
CogniPath is organized into five specialized microagents, each with a single responsibility:
Spatial-Audio Worker — receives an image, processes it with Amazon Bedrock (Claude 3.5 Sonnet), and generates four descriptions ordered by level of detail. Level 1 is general context. Level 2 is axis structure and scales. Level 3 is trends and relationships. Level 4 is the exact numerical data table. The student navigates with keyboard arrows or voice commands.
PictoManager — the new agent in this version. When a student taps a vocabulary term, PictoManager first searches ARASAAC via its public API. If the pictogram exists, it delivers it with alt-text and Polly audio. If it doesn't (which is the case for nearly all subject-specific STEM vocabulary), it calls Amazon Titan Image Generator with a master prompt that enforces strict rules: pure white background, maximum 2–3 flat colors, no text, central element occupying at least 70% of the space. Claude 3.5 Sonnet validates that the resulting pictogram actually represents the concept correctly before saving it. The second query for the same concept comes from DynamoDB cache: millisecond latency and zero inference cost.
Content Worker — reformats any text according to the student's profile. For dyslexia: Bionic Reading (the first half of each word in bold), chunks of no more than two sentences, cream background instead of pure white, comprehension checkpoints between fragments, clickable key vocabulary that triggers a narrated definition. For ADHD: one action at a time, no decorative elements competing for attention, progress auto-saved after every interaction.
Voice System — voice capture via Web Speech API (Amazon Transcribe Streaming in production), intent detection using Claude as a command classifier, response synthesis via Amazon Polly Neural. Voice selection is the first step of onboarding, not a setting buried in preferences. There are four voices: moderate feminine in Mexican Spanish, moderate masculine in Colombian Spanish, slow feminine for users who need more processing time, and low-pitch masculine. The same sentence plays for all four, so you can compare them directly.
Socratic Tutor — the tutor never gives direct answers. It always asks before confirming. Every response is anchored to a specific chapter of the official curriculum (Argentina 2024, Colombia, Mexico). If the tutor can't find the source, it says so explicitly instead of fabricating a citation. That decision — transparency about uncertainty — is what makes the tool trustworthy for teachers.
What's Working in the Prototype Right Now
The current prototype is a 2,450-line HTML file that demonstrates:
- Complete onboarding with no registration, three questions, under 60 seconds to first real use
- Spatial-Audio Worker with four keyboard- and voice-navigable levels
- Text with real-time Bionic Reading toggle, key vocabulary narrated on click
- Socratic tutor with conversation history and a "listen" button on every assistant response
- Visual pictogram agenda with session progress bar
- Narrated Pomodoro with verbal alerts at milestones (50%, 5 minutes remaining, completion)
- Voice command system with 20+ commands in Spanish, LATAM accent normalization, floating status panel
- Complete dark mode via CSS custom properties with smooth transition
- Narrator always on by default: the first time the app loads, it speaks before the user does anything
Voice synthesis uses the Web Speech API with the system's browser voices. In production, this gets replaced with Amazon Polly Neural and SSML for full prosodic control: pauses, pacing, emphasis on scientific keywords.
The AWS Production Stack
| Service | Role in CogniPath |
|---|---|
| Amazon Bedrock (Claude 3.5 Sonnet) | Multi-level description generation, pictogram validation, voice command classifier, Socratic tutor |
| Amazon Bedrock (Titan Image Generator) | STEM pictogram generation for vocabulary ARASAAC doesn't cover |
| Amazon Polly Neural | Voice synthesis with SSML — pause, speed, and emphasis control |
| Amazon Transcribe Streaming | Real-time voice recognition, <300ms latency |
| AWS Lambda | Five microagents, serverless architecture — zero cost during inactivity |
| Amazon DynamoDB | Pictogram cache, user profile, exact session position |
| Amazon S3 | Storage for generated pictograms and SSML audio |
| AWS Step Functions | Persistent Narrated Pomodoro — survives a locked screen |
| Amazon EventBridge Scheduler | Pomodoro milestone alerts without polling |
| Amazon Cognito | Anonymous user profile — no account, no barrier to entry |
| Amazon API Gateway | Streaming audio delivery to the React client |
The Design Decision That Changed Everything
The most important change I made wasn't technical. It was deciding that the narrator would be active by default instead of something the user enables.
The first version had an "Enable narrator" button on the welcome screen. The logical choice — what almost every accessibility app does. The problem: a user with a visual disability arrives at a screen they can't read, looks for a button they can't find because they don't know where it is, and leaves before hearing a single word.
The solution is having the app speak first. 800 milliseconds after loading, without the user doing anything, a voice says: "Welcome to CogniPath. To get started, tap anywhere on the screen or say 'start.'" From that moment, the user knows the voice interface works and can begin exploring.
That shift — from opt-in to opt-out for the narrator — summarizes CogniPath's design philosophy: accessibility isn't something you turn on. It's the default experience.
Demo
Screenshots to include:
- Welcome screen with active narrator (green indicator pulsing)
- Spatial-Audio Worker at Level 3 — trends in the temperature/enzyme activity graph
- Text screen with Bionic Reading active and vocabulary panel open
- Voice command panel with animated wave during listening
- Visual agenda with session pictograms and progress bar
- Full dark mode
- Promodoro
What I Learned
Voice is the interface, not an add-on. I originally designed CogniPath as a visual app with "audio support." After reading enough research on design for blind users, I understood that had it backwards. For a user with a visual disability, the screen is the add-on. Voice is the primary interface. That completely changes the order of design decisions: what happens before the user touches anything, what the system says when the user makes an error, how information is structured to be linear and predictable.
Silence is a feature. The decision not to play anything during the Pomodoro break was the one I second-guessed most. It feels like an omission. It's actually a deliberate decision based on research on cognitive recovery: rest only works if there's total disconnection from stimulus. Coding silence as functionality — not as the absence of code — changed how I think about AI system design.
Caching as an ethical principle. In the context of Latin American public schools, inference cost isn't just a business problem. It's an access barrier. Designing the pictogram flow with DynamoDB as the first stop, before calling Bedrock or Titan, wasn't a cost optimization — it was a decision about who gets to use the system. The second time anyone queries "mitochondria," the cost is zero. That zero has moral weight.
Discoverability in voice is hard and critical. I built the voice command system and the first informal test was a failure: nobody knew what to say. The solution wasn't documentation — it was having the app narrate its own capabilities when the user asks for help. The "help" command from any screen triggers a narration of the available commands in that context. I learned that in voice interfaces, discoverability has to be built into the flow, not written in a separate manual.
What's missing for the real version. The prototype uses the Web Speech API, which works well in Chrome and Edge but has inconsistent support in Safari and Firefox. In production, Amazon Transcribe Streaming fixes that and adds separate support for Argentine, Colombian, and Mexican Spanish accents. The VUI design research (AufaitUX, 2025; Designlab, 2024) is explicit about this: a recognition system trained on North American accents fails at an unacceptable rate for Latin American users. That's the difference between a demo and a tool someone uses every day.
The biggest thing I learned from this challenge: building for accessibility isn't about adding features for "special users." It's about examining from the beginning what the standard design assumes and who those assumptions leave out. Every decision I made for accessibility ended up improving the experience for everyone.
CogniPath — The barrier isn't intellectual. It's the format.
AWS Cloud Club Captain UCASAL Misael Calisaya · Salta, Argentina · March 2026
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article