Urban — About the Project

What inspired us

Urban was inspired by a simple but frustrating reality: city policies are ambitious on paper, but the path from a citizen issue → policy rule → real-world action → measured impact is slow, fragmented, and hard to track. We wanted to build something that feels like a “government consulting team in a box” — an agentic system that can understand policy, ingest real-time signals, recommend compliant actions, and show measurable progress. The SF Urban Housing & Development Policy (SFUHDP-2030) we analyzed became our perfect sandbox because it has clear targets, timelines, and accountability mechanisms.

What Urban does

Urban is a privacy-first, production-grade civic agent platform for city teams. It:

Triages citizen and staff cases (housing, infrastructure, safety, environment, funding).
Retrieves relevant policy clauses automatically and cites them in responses.
Pulls live context from the web/open data sources for evidence-backed recommendations.
Generates department-specific action plans + citizen-safe response drafts.
Simulates policy outcomes to show likely post-effects of decisions.
Tracks progress against targets in a real-time analytics dashboard.
Visualizes impact on a live 3D city map, with layers for zoning, projects, and environmental upgrades.

Key features we built

1) Custom agents (consulting-style)

We designed multiple focused agents that collaborate:

Policy Analyst Agent: reads policy docs, extracts targets/milestones, answers clause-level questions.
Case Triage Agent: classifies incoming issues, ranks urgency, routes to the right department.
Evidence Scout Agent: fetches real-time web/open-data signals and summarizes them.
Impact Simulator Agent: runs quick “what-if” projections for targets and side-effects.
Risk/Compliance Agent: checks for sensitive cases or financial/crypto risks before action.

2) Policy simulations

Instead of a static dashboard, we built simulation controls that let officials explore tradeoffs:

Density vs affordability effects
Budget allocation vs pace of completion
Infrastructure capacity vs rezoning plans
We modeled simple projections such as: [ \text{progress\%}=\frac{\text{current units built}}{95{,}000}\times 100 ] and pace requirements like: [ \text{units needed per year}=\frac{95{,}000-\text{built so far}}{2036-\text{current year}} ]

3) Analytics dashboard (policy-aligned KPIs)

The analytics page mirrors SFUHDP’s actual monitoring logic. We show:

Housing production vs target (95,000 by 2036)
Affordable housing vs 40% requirement (~38,000 units)
Commute time change vs −15% target
Emissions change vs −22% target
Housing cost burden vs ≤30% target
Roof greening/solar adoption vs 30% by 2033 + 1.2M sq ft by 2036
Modular Housing Fund efficiency
$5.4B budget burn-down and funding mix vs plan
Agency scorecards and backlog/SLA health
This makes Urban feel like a real consulting PMO tool, not a demo toy.

4) Live 3D map of post-effects

To make policy outcomes tangible, we added a 3D city view (prototype layer):

Pins for projects by status (planned / approved / under construction / complete).
Heatmaps for affordability gaps and commute pressure.
Environmental overlays for cooling corridors and green roof spread.
This “post-effects view” is the storytelling core: officials can see how choices reshape the city.

5) Privacy-first design

Because civic data includes sensitive PII, we built a strict boundary:

PII stored securely and tokenized before any LLM call.
Only minimal, redacted summaries reach the reasoning layer.
Audit-friendly logs for compliance.

How we built it

Our build was a fast, hackathon-grade version of a production system:

Document understanding: extracted and structured policy text into targets, rules, timelines.
Knowledge modeling: stored policy + programs + milestones in a structured CMS.
Retrieval + memory: indexed policies and past case summaries for fast semantic lookup.
Agent orchestration: connected tool-calling agents with clear JSON schemas.
Simulation layer: simple forecasting formulas + interactive UI sliders.
Dashboard UI: KPI cards, charts, scorecards, and a 3D map layer for impact.
End-to-end demo flows: crisis case → policy lookup → evidence → action plan → analytics update.

What we learned

Policies become powerful only when they are operationalized. We learned how to convert dense government PDFs into checkable rules + KPIs.
Agentic systems need guardrails. Civic agents must be safe, minimal-question, privacy-aware, and auditable.
Dashboards without simulations are blind. Real decision tools need “what-if” views, not just reporting.
The story matters. The 3D post-effects layer dramatically improved how intuitive policy impact felt.

Challenges we faced

PDF complexity → structured truth: policies are long, verbose, and sometimes inconsistent; extracting reliable targets was non-trivial.
Simulation realism vs time: we had to keep models simple but believable in one day.
Tool coordination: ensuring agents worked with retrieval, evidence tools, and privacy layers without breaking the flow.
Avoiding “dashboard spam”: we focused on policy-aligned metrics only, not random charts.
Civic safety: ensuring the system doesn’t give unsafe advice, especially in sensitive cases.