Immi AI

Dashboard
Live Map
Architecture Diagram

Inspiration

Undocumented immigrants face sudden ICE raids, language barriers, and limited access to reliable legal guidance—often on mobile only. We set out to create a lightweight, mobile-first service that empowers vulnerable communities with real-time situational awareness and know-your-rights support, whether via chat, call, or a simple panic button.

What it does

Live Map of ICE Activity: Continuously scraped from community and trusted sources, plotted on an interactive map with mobile support.

Multilingual RAG Chatbot: Spanish-first chatbot powered by Pinecone retrieval over legal docs; falls back to lawyer escalation on low-confidence answers.

Integrated Voice Agent: Phone-based voice interface (via Vapi) for users without internet: menu prompts, speech-to-text Q&A, real-time translation. It also includes an option to call a number that connects users to a Vapi agent for live conversation.

Panic Mode: One-tap record-and-send feature to instantly alert a designated emergency contact with a video recording and location.

Legal Document Translator: Interface to scan and take pictures of legal pdfs, translate, and explain legal paperwork on the fly.

How we built it

Frontend: Next.js on Vercel, React-Leaflet for the map, Tailwind CSS for responsive layouts.

Backend: FastAPI + Uvicorn running on EC2; MongoDB for user/session data; Firecrawl for ICE-activity scraping pipelines.

AI Services: Pinecone vector store for retrieval-augmented generation over curated legal documents; Vapi + Trieve for telephony and voice-agent flows.

Challenges we ran into

Data Reliability: Building a robust scraper that filters noise and false-positives in community reports.

Latency vs. Recall: Tuning RAG source counts to balance speed on low-power devices with answer accuracy.

Connectivity Constraints: Ensuring smooth UX on poor cellular networks and seamless fallback between SMS, chat, and voice.

Ensuring Credibility: Guaranteeing that sources are credible and transparent.

Accomplishments that we're proud of

Fully mobile-optimized platform with zero-install voice and SMS support.

Spanish-first chatbot delivering legally grounded answers rapidly.

Real-time location-aware map updates with functionality to report events, helping users reroute away from active raids.

Panic Mode feature to send an emergency recording to your emergency contact's email.

What we learned

Real-time legal support demands tight integration of scraping, RAG, and telephony.

Multimodal pipelines (chat, voice, camera) can coexist smoothly on mobile with careful resource management.

Community-driven data needs rigorous validation to be trustworthy.

A minimal UI combined with powerful backend tooling can deliver complex features to low-connectivity users.

What’s next for Immi AI

We are super excited and motivated by this project! Some of the features we would love to make further are: Roll out domain, branding (logo, site icon), and production SSL.

Extend language support beyond Spanish/English.

Partner with legal aid organizations for live lawyer on-ramps.

Build analytics dashboard for usage insights and data quality monitoring.

Built With

ec2
fastapi
firecrawl
gemini
javascript
mongodb
next.js
openai
pinecone
python
trieve
uvicorn
v0.dev
vapi

Submitted to

UC Berkeley AI Hackathon 2025

Created by

I designed the app frontend and fullstack RAG chatbot with Vercel and Next.js. I helped upload sources to the Pinecone vectordb. I worked on mobile support, photographing documents, and the emergency email panic feature. I also worked on polishing the UI and deploying to AWS/Vercel.

Alexander Anokhin
I created a system to monitor and visualize real-time ICE raid by utilizing a web crawler that gets location data from publicly available sources. Leveraging FastAPI, I developed a backend that stores and serves this data through a RESTful API, powering an interactive map that updates dynamically every few hours. I implemented geolocation features to place raid pins on the map and built a search functionality that enables users to find city-specific data quickly. Users can also contribute crowd-sourced reports via a submission form, which are integrated into the live dataset. To extend accessibility, I integrated Vapi to deploy a voice-based legal assistant, accessible directly from the website or by phone, that provides multilingual legal guidance through natural language conversation.

Anish Nambirajan
I handled MongoDB integration and API requests to manage user sessions and ICE activity data. For the map, I contributed to scraping ICE activity from Padlet and other online sources, and built an interface to support additional crowdsourced reports. I also implemented location sensing to detect nearby ICE activity and seamlessly route users to our RAG-powered, immigration-specialized chatbot for real-time legal support if necessary.

cyril Joby
I developed a custom web scraper to extract legal text from Title 8 of the Code of Federal Regulations. Then, I built a data processing pipeline to chunk text, vectorize chunks, and upload chunks to Pinecone for downstream RAG functionality. Furthermore, I implemented a legal document translation pipeline that integrates pdf processing and Gemini for multilingual support. I also built out an intuitive map framework that supports marker creation and global visibility.

Rishi Murumkar