I'm 13 and built a private AI legal assistant for lawyers — here's what I learned

I'm 13, and for the past few months I've been building Vaultr — a private AI legal assistant for lawyers. Here's what I've learned.
The problem I set out to solve
Lawyers using ChatGPT or Harvey risk losing attorney-client privilege. A US court recently ruled that lawyer-AI conversations lost privilege because the provider could retain and share data. That's a massive problem. Vaultr solves it architecturally — by never sending data anywhere in the first place.
What I built
Vaultr runs on Next.js 16 + Tauri v2 for macOS. It gives lawyers two modes:
Private mode — 100% local via Ollama. Nothing leaves your machine. Models: Phi-4 Mini, Qwen3 8B, Gemma.
Cloud mode — more powerful models when needed (Llama 3.3 70B, Qwen3 32B via Groq, Gemini 2.5 Flash). Lawyer chooses.
Features: contract risk scanning, auto redaction, AI legal assistant (Lex), web search for real-time legal info.
What I learned
Balancing school and building is genuinely hard. There were nights debugging Tauri builds before exams. But shipping something real — even imperfect — teaches you more than any classroom.
The biggest lesson: talk to your users before building everything. I almost spent months building a massive legal database before realising I should validate the core product first.
What's next
.dmg build, beta launch with 10 lawyers, then iterate.
If you're a lawyer or know one — waitlist here: tally.so/r/ODkzjk
Would love feedback from this community! 🚀