Declaro

Inspiration

Earlier this year I watched a PBS documentary about children being separated from their parents at the border. Around the same time, ICE raids started happening in my own hometown in New Jersey. People in my community were terrified, not because they were criminals, but because they had no idea what their rights were. Less than 1% of people in low-income countries have access to legal aid. 123 million people are displaced with no way to get help. I built Declaro to help people get legal aid, without worrying about safety and money.

What It Does

Declaro is a voice-first multilingual legal aid assistant. Speak your situation in any language, hear your legal rights back in under 3 seconds. No reading required. It supports 65+ languages, detects emergencies like domestic violence and detention, surfaces real hotline numbers, and generates formal legal documents tailored to your specific situation and location.

How We Built It

Declaro runs on a three-API pipeline: OpenAI Whisper for speech recognition, Anthropic Claude Sonnet for legal reasoning, and OpenAI TTS for voice synthesis. On top of that I built a real-time Google Translate integration across 65 languages, a confidence scoring system, multilingual emergency detection, and a context-aware document generation engine that writes in the official language of the user's country. Deployed on Vercel with zero data storage.

Challenges We Ran Into

MyMemory API was mapping Spanish to Esperanto and Gujarati to Punjabi, so I switched to Google Translate and fixed it. Whisper was hallucinating random text from silent recordings, which I fixed by rejecting audio blobs under 10,000 bytes. Claude kept responding in Hindi when users selected Gujarati, so I fixed it by passing explicit script instructions with every prompt.

Accomplishments That We're Proud Of

Getting the full voice pipeline working in under 3 seconds across 65 languages. Building a document engine that knows the official legal language of any country and auto-appends a translation in the user's spoken language. And doing all of this in 5 days at 15 years old.

What We Learned

Building for accessibility is completely different from building for convenience. Every decision had to account for users who might be scared, illiterate, or on a cracked phone in a detention center. Language is also genuinely hard, the scripts, the directionality, and the way AI models confuse similar languages. Getting that right took more engineering than anything else in the project.

What's Next for Declaro

Partnering with refugee organizations like UNHCR and RAICES to get Declaro into the hands of people who need it most. Adding more location-specific legal frameworks, offline support for areas with poor connectivity, and a verified legal review layer so every response is checked against real case law.

Built With

• anthropic-claude
• google-translate-api
• next.js
• openai-tts
• openai-whisper
• typescript
• vercel