AI-mmunity

Inspiration

One of our teammates has a child with respiratory issues. Figuring out “Is today safe to be outside?” is a daily hassle made harder because the most important info (air quality, disease trends, local resources) is scattered across government datasets and public repositories. We realized many vulnerable groups—kids, seniors, people with chronic conditions, outdoor workers—face the same barrier. We set out to build a single, scalable place that turns raw, siloed data into practical, hyper-local guidance, and to give public health officials the granular signals they need to spot issues earlier.

What it does

AI-munity is a serverless, multi-agent platform that delivers conversational, hyper-local health intelligence for both residents and public health officials:

For residents: Personalized, county-level risk insights (e.g., air quality, disease signals) plus clear, actionable recommendations available in multiple languages with speech in/out so everyone can use it.

For officials: A crowdsourced health signal + analytics layer with real-time, agent-driven dashboards, alerts, and synthesized reports to guide resource allocation (e.g., mobile clinics) and detect emerging issues earlier.

Core features: Multi-agent chat (Air Quality, Disease, Health FAQ), alerts & recommendations, translation + TTS/STT, and interactive visualizations bundled into a UI/UX-friendly web app.

How we built it

Architecture: Deployed on Cloud Run with a Flask API exposing /api/air-quality, /api/health-recommendations, /api/agent-chat (ADK multi-agent), and /api/translation.

Multi-agent system (ADK): Air Quality Agent, Disease Agent, and Health FAQ Agent orchestrate data lookups and summarize guidance, Veo Agent for PSA's and Social Medial agent to post to twitter.

Data & AI:

BigQuery with real EPA air quality data (2010–2021); additional CDC/WHO disease datasets.

Gemini for natural-language reasoning and synthesis.

Frontend: Tailwind CSS, Three.js (animated background), Chart.js & D3 for visualizations, responsive mobile-first design, location selector, loading and error states.

Data engineering: Efficiently handle 241K+ rows, pandas/numpy aggregation, robust date parsing, missing-value handling, and low-memory CSV parsing.

UX: Agents are embedded directly into the app so users get insights, visuals, and follow-ups in one place.

Challenges we ran into

Data fragmentation & quality: Normalizing formats and semantics across EPA/CDC/WHO sources and dealing with gaps/missing values.

Hyper-local context: Mapping coarse datasets into actionable county-level guidance without over- or under-stating risk.

Real-time feel on serverless: Balancing cold starts, caching, and responsiveness for conversational experiences on Cloud Run.

Multi-agent coordination: Getting agents to share context cleanly and avoid redundant or conflicting answers.

Accessibility: Designing for multi-language support and speech interfaces so vulnerable users can actually use the tool.

Accomplishments that we're proud of

A working end-to-end serverless app that merges multi-agent chat with interactive dashboards.

County-level insights and personalized recommendations synthesized from large public datasets.

Translation + speech (TTS/STT) so guidance is accessible beyond English and text-only UIs.

Clean, modern data visualizations and a UX that makes technical data understandable.

Handling 241K+ rows smoothly with solid data-wrangling and performance optimizations.

Twitter integration and semantic understanding of images uploaded by users, to assess severity of the report

What we learned

Human-centered AI > data dumps: People need “what it means and what to do,” not charts alone.

Serverless + agents pair well for rapid iteration, but you need caching and careful prompt/data design for snappy responses.

Data governance matters: Provenance, explainability, and conservative wording are key for health guidance.

Crowdsourcing is powerful—but incentives, trust, and moderation policies are as important as the tech.

What's next for AI-mmunity

We reached to educators with-in UCSF medical school to evaluate our platform and learn how we can enhance its offering. They were very impressed with how intuitive it was and all the features it has out of the box. We're now if a pilot is possible with-in their population and community od medical faculties with-in the UC system.

we're also working on

Live feeds & richer signals: Ingest near-real-time EPA AQI, CDC nowcasts, wastewater and syndromic data; expand to heat, pollen, and wildfire smoke.

Deeper official tooling: Case management views, hotspot detection, exportable summaries, and alert workflows for city/county agencies.

Stronger community reporting: Lightweight, privacy-respecting reports with incentives and verification to sharpen local signals.

Access & trust: Mobile app, push notifications, more languages, offline-first patterns, and clear model cards & disclaimers.

Partnerships & scale: Pilot with schools, employers, and local health departments; instrument evaluation dashboards and A/B tests.

There's only so much we can put into a demo video, we implore you to explore all the feature on our site, nothing is static all of the features, data and outputs are real wither from our agents or many datasets we have.

(the cold start dose takes about a min to load the first time you hit the URL so please be patient but after that it should work as expected)