An open-source, adaptive AI writing tool built for dyslexic thinkers.
Dyslexia isn't a deficit — it's a different cognitive architecture. DysLex AI removes the friction between brilliant thinking and the page.
- Per-user adaptive error profiling — learns your unique spelling and grammar patterns over time
- Instant local corrections — small ONNX model runs in-browser, under 50ms, works offline
- Deep contextual analysis — catches homophones, real-word errors, and severe misspellings via NVIDIA Nemotron
- Passive learning — no accept/reject buttons; the system learns silently from your natural writing behavior
- Voice-first input — real-time transcription with idea extraction via faster-whisper
- Text-to-speech — hear your writing read back to you via MagpieTTS
- Vision input — upload handwritten notes, diagrams, or whiteboard photos
- Four writing modes — Capture (voice-to-thought), Mind Map, Draft (scaffolded writing), Polish (review and refine)
- Dyslexia-friendly design — specialized fonts, eye-friendly colors, invisible corrections, no deficit framing
# One command to start everything
python3 run.py --auto-setupThis starts PostgreSQL, Redis, the backend (port 8000), and frontend (port 3000) automatically.
# Or run manually
cd frontend && npm install && npm run dev # port 3000
cd backend && pip install -r requirements.txt && uvicorn app.main:app --reload # port 8000
# Or use Docker
docker compose -f docker/docker-compose.yml upPrerequisites: Python 3.11+, Node.js 20+, PostgreSQL 15+, Redis — see Quickstart Guide for details.
| Model | What It Does | Where It Runs |
|---|---|---|
| Error Profile | Learns each user's unique error patterns over time | PostgreSQL (server) |
| Quick Correction | Instant fixes for known errors — no network needed | ONNX Runtime (browser) |
| Deep Analysis | Context-aware correction for homophones, real-word errors, intent | NVIDIA Nemotron (cloud) |
These three models feed into each other through a passive learning loop — the system observes natural writing behavior without ever asking the user to accept or reject anything. See How It Works for the full architecture.
Frontend: React 18, TipTap, TypeScript, Vite, Zustand, ONNX Runtime Web Backend: Python 3.11+, FastAPI, SQLAlchemy 2.0, Alembic, Pydantic 2.0 ML/AI: NVIDIA NIM (Nemotron), MagpieTTS, faster-whisper, ONNX Runtime
- Quickstart Guide — Prerequisites, setup, environment variables, troubleshooting
- How It Works — Architecture, passive learning, writing modes, accessibility
- Testing — ML tests, frontend/backend test suites, benchmarks
- Architecture — System design overview
- API Reference — Endpoint documentation
- Database Schema — ERD and table docs
- ML Models — Model architecture and training
- Deployment — Production setup
- Contributing — How to get involved
Apache 2.0 — See LICENSE
Dyslexia affects roughly 1 in 5 people worldwide. A tool this powerful should not be locked behind a paywall.
"Everyone has ideas worth sharing. The right tool just makes sharing them easier."
