ADDY

TAPIA / Capital One 2026 — challenge [2] Generative AI Neurodiversity-Adaptive Study Workspace.

GitHub: aidaken/xforce-hack

Product name: ADDY. One passage in, two remakes out — ADHD checklist vs dyslexia flowchart — then a source check so we did not invent or drop anything. Toggle the learner mode in the workspace. Student taps This isn’t working → the agent switches format.

What is actually shipped (API, limits, gaps): product_info.md. Agents update that file after every finished task. This README is the pitch / MVP draft.

Direction is drafted here; a few scope questions are still open.

Why this is necessary

Dense textbooks and technical docs are written as long paragraphs. That format is a poor fit for ADHD (working-memory load, no “what do I do next”) and for dyslexia (decoding load, walls of justified text). Existing tools summarize or restyle the same prose. Judges asked for restructuring: flowcharts, checklists, or gamified summaries, tailored to the learner, with an agent that chooses the format, checks itself against the source, and adapts when the student says it failed.

We are not diagnosing anyone. Copy in the app: designed around common ADHD / dyslexia study preferences, not clinically proven.

What the app looks like

Landing. Short pitch + Remix a passage (demo loads a preloaded OpenStax passage immediately). Optional later: sign in for a saved profile.

Workspace (the product). One screen.

• Left: source passage (the teammate’s human-check surface).
• Right: remake for the active learner (checklist, Mermaid flowchart, or gamified step-reveal).
• Top: learner mode / preference toggles, planner rationale (“I chose a flowchart because this is a 3-stage process…”).
• Bottom / panel: fidelity report (supported / missing / invented) with source spans and linked highlighting.
• Adapt: This isn’t working + a reason → re-plan to a different format or granularity.

Demo never depends on live copy-paste: three dataset passages are preloaded.

Demo we will show judges

Same passage, two different learners, human side-by-side check against the source.

1. Load Cellular Respiration (or another preloaded source).
2. ADHD remake: checklist / one-step-at-a-time.
3. Toggle to dyslexia remake: large-type flowchart (OpenDyslexic or Atkinson Hyperlegible, short labels).
4. Open the fidelity panel. Source Guard teammate confirms nothing was lost or invented.
5. Tap This isn’t working → agent switches format live.

Tight ship slice if time is short: one screen, one OpenStax passage, three buttons. No accounts, no uploads.

Features (MVP)

Feature	What it is
Login / profiles	Saved learner profiles. Demo slice may skip this.
Preference profile	Color theme, music, fonts, interests, 3–5 toggles, plus free-text “what trips me up.”
Input	Paste text, PDF (including scans), Word .docx, Google Doc (public link or .gdoc), or URL. Always preload the three challenge passages so the demo cannot fail on paste.
Planner agent	Classifies the passage (sequential process / decision logic / definition-heavy), picks a format, shows a one-line rationale. That line is what proves it is agentic, not a template.
Renderers	Flowchart (Mermaid), checklist, gamified summary (step reveal + quick-check questions).
Fidelity checker	Decompose source into atomic claims; map each output element to a source span. Flag unsupported (invented) and uncovered (lost). Show as a report.
Side-by-side	Source left, output right, linked highlighting. This is the human-check tool and the judge artifact.
Adapt loop	This isn’t working + reason → re-plan to a different format or granularity.
Lock screen	Optional focus lock for a student-set duration. Feasibility TBD.

Do not claim clinical diagnosis. Do not build a second unrelated product.

Fonts

ADHD-oriented (reduce crowding / re-reading) — accessibilitychecker.org

Font	Why	Use
Lexend	Built to reduce visual crowding; fluency research	Body, long-form remake text
Atkinson Hyperlegible	Distinct characters, less re-reading	Body + UI
Verdana	Wide spacing, large x-height, universal	Body, forms
Open Sans	Clean, uncrowded, even widths	Web body
Arial / Helvetica	Low visual noise	Fallback
Georgia	Screen serif, wide tracking	Long-form / print if needed

Dyslexia-oriented

Font	Why	Use
OpenDyslexic	Weighted bottoms; letters less likely to flip — opendyslexic.org	Dyslexia remake body
Atkinson Hyperlegible	High character distinction without a “special” look	UI + captions

Default pairing unless we change it: ADHD remake in Lexend, dyslexia remake in OpenDyslexic, chrome/UI in Atkinson Hyperlegible.

Preloaded demo sources

From the challenge prompt. Ship with all three in a dropdown; live demo defaults to biology.

1. OpenStax Biology — Cellular Respiration (3 dense paragraphs)
2. Technical docs — AWS IAM policy evaluation or React useEffect
3. Dense STEM — OpenStax Calculus §3.6 Chain Rule or University Physics §6.3 Gauss’s Law

Shared Supabase (all branches)

There is one hosted project. Use it on main, dev, aidar-be, and every other branch. Do not spin up a second project.

Name	xforce-hack
Dashboard	https://supabase.com/dashboard/project/tfmzjvoqsktlzwnrzdzr
Ref	tfmzjvoqsktlzwnrzdzr
API	https://tfmzjvoqsktlzwnrzdzr.supabase.co

After clone:

cp .env.example .env
npx supabase link --project-ref tfmzjvoqsktlzwnrzdzr
npm install
node scripts/check-supabase.mjs

Teammate guide: docs/supabase.md
Pinned IDs: config/supabase.json

Shared Vercel (all branches)

There is one Vercel project. GitHub aidaken/xforce-hack is already linked. Do not spin up a second project.

Name	xforce-hack
Dashboard	https://vercel.com/aidars-projects-c6143ce8/xforce-hack
Project ID	prj_GXZ75tI8NCSyxSXNFiHanqcIp1Zc
Team	team_XS9LEdNQcwp4cYraijcCL30b
Production	https://xforce-hack.vercel.app

After clone:

npx vercel link --yes --project xforce-hack --scope aidars-projects-c6143ce8

Teammate guide: docs/vercel.md
Pinned IDs: config/vercel.json

Pushes to GitHub create deployments. Feature branches get preview URLs. Aidar’s “push to prod” updates GitHub main and Vercel production. Put Supabase anon / NEXT_PUBLIC_* keys on this Vercel project; never the service-role key. Put OPENROUTER_API_KEY on Vercel too — never in git.

Frontend: docs/frontend.md — React SPA in app/ (Vite). Backend: server/ + api/.

Local:

cp .env.example .env   # add OPENROUTER_API_KEY locally, never commit .env
cp secrets.toml.example secrets.toml
npm install
npm run build
npm run dev

Then open http://localhost:3000/app.

Repo map

• product_info.md — live product record (update after every task)
• ideas/ — earlier pick-list (formats, agent loop, judge demo)
• app/ — React SPA (Vite). Edit here. Builds to gitignored public/app.html
• public/study-activities/ — Charlotte's portable guided reading + play widgets (wired into /app after ingest)
• public/index.html — landing page (plain HTML)
• server/ — ingest, classify, OpenRouter
• api/ — Vercel serverless wrappers
• docs/frontend.md — where frontend work starts
• supabase/ — CLI config, linked to the hosted project above
• CLAUDE.md / CURSOR.md / AGENTS.md — agent conventions; product truth is product_info.md

Do not checkout main to edit files. Feature work happens on branches. Aidar’s “push to prod” updates GitHub main and Vercel production.

Open decisions

Not locked yet. Highest-tension items:

1. Demo slice vs full MVP — no accounts / no uploads vs login, PDFs, URLs, lock screen.
2. Two named modes (ADHD vs dyslexia) vs preference toggles + free text as the primary learner model.
3. Planner picks one of three renderers vs always show both ADHD checklist and dyslexia flowchart for the judge toggle.

Full question list is in the current planning thread; answers should land in product_info.md (and here if the pitch changes).