Inspiration

AI City Council is a hackathon project that builds a real-time multi-agent AI system to empower San Francisco-based NGOs with better civic advocacy tools.Inspiration San Francisco NGOs often struggle to quickly turn vast amounts of open government data into targeted, evidence-based advocacy. We wanted to create always-on AI "city council members" that continuously watch official data and propose actionable solutions, bridging the gap between data and real policy change.

What it does

4–6 specialized AI agents (representing departments like Budget, Planning, SFPD, Public Works, etc.) run in marathon mode: They pull live data from data.sfgov.org every ~60 seconds
Analyze for issues (e.g., budget shortfalls, rising crime trends, service backlogs)
Generate city-wide policy recommendations with cited evidence
Publish issue + solution cards to a live dashboard Agents collaborate when topics overlap (e.g., Budget + Planning). NGOs click any card to trigger a Composio-powered agent that drafts professional, citation-rich advocacy emails to the right city officials.

How we built it

Backend: Python + FastAPI, LangChain for agent orchestration, Socrata API for SF open data, You.com API for context & precedents
Agents: Custom classes running continuous 60-second loops for fetch → analyze → solve → publish
Collaboration: Agents monitor each other's outputs to detect & produce joint proposals
Email drafting: Composio handles structured NGO input → professional email generation
Frontend: React/Next.js dashboard with real-time updates (polling/SSE), Tailwind styling, severity-coded cards
Deployment: Vercel (frontend) + Railway/Render-like (backend)

Challenges we ran into

Managing API rate limits and data freshness on data.sfgov.org
Designing reliable cross-agent collaboration detection without excessive compute
Balancing agent analysis speed (target <30s per loop) with meaningful insights
Crafting consistent output schemas and realistic official contact mapping
Keeping email drafts professional, specific, and minimally editable while injecting real citations

Accomplishments that we're proud of

Got 4–6 agents running continuously with live dashboard updates during the demo
Demonstrated meaningful cross-agent collaboration (joint cards appeared)
Composio agent successfully generated fully-cited, actionable email drafts addressed to real SF officials
Created specific, evidence-based policy proposals rather than generic suggestions

What we learned

Multi-agent systems shine for complex, intersecting civic problems but require careful orchestration to avoid noise/loops
Open data portals like Socrata are powerful but need smart caching & change detection
NGOs value specificity (exact recipients, numbered asks, timelines) far more than vague AI enthusiasm
Real-time + marathon agents create a very different feel from one-shot chat-style AI

What's next

for AI City Council Expand to more agents and deeper historical trend analysis
Add lightweight predictive signals (e.g., early warning thresholds)
Integrate actual NGO feedback loops and real email send (post-hackathon)
Explore district-level granularity and mobile-friendly views
Open-source core agent framework for other cities to adapt

Built With

• antigravity
• claude
• composio
• javascript
• python
• youdotcom