Gaurdian

Inspiration

I'm heading to law school next month, and before I left, I wanted to build something for the community that could actually make a difference in the field I'm about to enter. When I heard about the USAII Hackathon, it felt like the right place to do that — a chance to help in my own field, law, using AI. When the challenge briefs were revealed, Guardian was the one I knew I had to build.

Guardian is for mandated reporters — people who have to step in when they believe a child is in trouble, often in the most pressured, uncertain moments, when they don't know what to do or what not to do. A child dies from abuse every three days in America. Most of the time, it's not because the case went unreported — it's because it was reported incorrectly, or the reporter didn't know who to report it to. That's a solvable problem, and it's the one our team set out to solve.

What it does

Guardian takes a mandated reporter's description of a situation in plain English and gives them a risk classification, the exact statute that applies, who to contact, and what to do next currently across five states.

How we built it

First, I built out a knowledge base in Obsidian, organizing the statutes and legal codes for all five states — this became our RAG knowledge base. On top of that, we run a two-model pipeline on Groq: Llama 3.3 70B for legal reasoning against that knowledge base, and Llama 4 Scout for reading photo evidence.

For the rest of the stack: Next.js with Tailwind for the frontend, Clerk for authentication, Upstash for rate limiting, and Vercel to deploy. We used Claude Code and VS Code to build it.

Challenges we ran into

The first real challenge was accuracy. Two of my three teammates didn't come from a legal background, and I wasn't fully there myself yet either, so every piece of legal information the app gave out had to be fact-checked by hand before we trusted it. Making sure Guardian never hallucinated a citation was a hard requirement from day one, not an afterthought.

The second challenge came once the challenge briefs were revealed we realized work done before the brief was announced might not actually fit what was being asked for, so we had to go back, strip parts of what we'd built, and rebuild to target the brief precisely instead of our original assumptions.

The third, and honestly the biggest, challenge was that this was my first time ever building an app. I come from a non-technical background, so every tool Obsidian, Claude, VS Code was new to me. Figuring out how all of it worked, and fixing bugs that probably would've taken a technical person minutes instead of hours, was the hardest part of this whole process.

Accomplishments that we're proud of

Getting through all of that learning every tool from scratch while still shipping something real while a non-technical person, is what I'm proudest of. It was difficult, but genuinely exciting to learn while building something that could actually matter.

What we learned

Building Guardian taught me how powerful and how useful AI can actually be in the legal system, and how much work this specific field still needs. Child abuse isn't just a US problem, and underreporting is a serious problem in its own right, everywhere. Beyond the technical side, this process taught me how to use AI more effectively and more ethically, and it gave me a wider view of the problem than I had going in. We've already started reaching out to school counselors and NGOs who want to help.

What's next for Guardian

We want to expand past our current five states to cover all fifty, and eventually to other countries as well. We're actively gathering feedback by reaching out to NGOs, schools, and hospitals anyone who needs this and has been navigating this kind of confusion without a tool like it.

Built With

• 14
• app:
• authentication:
• case
• claude
• clerk
• code
• css
• deployment:
• dev
• fl
• frontend
• il)
• limiting
• models-/-ai:-llama-3.3-70b-?-legal-reasoning-agent-(against-rag-knowledge-base)-llama-4-scout-?-vision-model
• next.js
• ny
• rate
• reads-photo-evidence-groq-api-?-inference-for-both-models-critic-agent-?-separate-citation-verification-step-before-any-statute-is-shown-knowledge-base:-obsidian-?-used-to-build-and-organize-the-rag-knowledge-base-(statutes/codes-for-ca
• redis
• storage:
• tailwind
• tools:
• tracker
• tx
• upstash
• vercel
• vs