Inspiration
"Loquere lingua tua. Omnes te audient sua lingua. Tua voce." (Speak your language. Everyone hears you in theirs. In your voice.)
In 1348, Charles IV founded the first university in Central Europe, uniting nations through knowledge and language. In 2026, his AI namesake finishes what he started — eliminating every communication barrier, including for those who cannot hear or see.
The world has 7,000 languages, 466 million deaf people, and 285 million visually impaired. We built one system that serves them all.
## What it does
Web4Light is a formally verified AI ecosystem powered by Gemini:
- Karel IV. — Universal translator with AI dubbing. Reads subtitles, translates, and speaks in your voice. Real-time, any language pair.
- Mozart — Sign language agent for the deaf. Converts any audio/text stream into sign language via AI avatar. Free for all deaf users.
- Žižka — Audio navigation for the blind. Camera feed → Gemini Vision → spoken guidance. Free for all blind users.
- Hus — The brutally honest AI assistant. Formally verified to never hallucinate.
- Asgard Studio — AI agent factory with SPARK certification (MUSH DO++ testing).
- Mincovna (The Mint) — Transparent pricing. Standard 700: 1 coin = 12g silver.
## How we built it
The core is written in Ada/SPARK — a language used in aviation and defense where failure is not an option. Every function has mathematically proven contracts:
$$\forall x \in Input: precondition(x) \implies postcondition(f(x))$$
Zero runtime errors. Not by testing — by proof.
Architecture:
- Whisper (STT) → Gemini API (translation/reasoning) → Spark-TTS (voice cloning)
- SPARK formal verification guarantees pipeline correctness
- Solar-powered self-hosted servers (zero operational cost)
- Google Cloud + Gemini API for inference
- Groq for ultra-low latency responses
AI-Native Operations:
- AI agents handle translation, dubbing, sign language generation autonomously
- SPARK prover runs continuously — if verification fails, the pipeline halts (fail-safe)
- No human in the loop for standard operations
## Challenges we ran into
- No AI frameworks exist for Ada/SPARK. Every component was built from scratch with formal verification contracts. We are teaching GNAT Czech.
- Voice cloning with timing constraints. Dubbing must finish before the next subtitle. SPARK proves timing mathematically.
- Sign language is not word-for-word. It has its own grammar. Gemini handles transformation, SPARK verifies completeness.
## Accomplishments that we're proud of
- Full pipeline working: YouTube URL → translate → dub → sign language in 39 seconds
- 61 subtitles translated at 100% accuracy via Groq fallback
- Zero operational cost (solar-powered)
- Live website: web4light.online
- Built entirely in one night by one person with AI assistance
## What we learned
A single person with zero programming background + AI assistants can build what used to require a 50-person engineering team. In one night. Powered by the sun.
Also: GNAT accepts Czech identifiers. The compiler does not judge.
## What's next for web4light
- Full AI avatar sign language generation (Mozart)
- Žižka deployment with smart glasses
- 10 parallel SPARK instances → 160 languages formally verified
- Open-source SPARK bindings for Gemini API (first in the world)
- Swiss AG incorporation for international scaling


Log in or sign up for Devpost to join the conversation.