Superblock

Real-time infrastructure for human-centered, health-aware city planning

Comment

Inspiration

Our planet's future hangs in the balance, and nowhere is the friction more visible than in our cities, where "gray infrastructure" accelerates climate change. Urban heat islands kill 1,300+ Americans annually, disproportionately affecting low-income communities who lack the clean energy resources to adapt. Traditional planning is too slow for a planet in crisis; it relies on cold satellite imagery that misses the human cost. We asked: What if we could build a living, breathing urban nervous system to sustain the spark of our communities by sensing environmental injustice in real-time?

What it does

Superblock is a decentralized urban nervous system that turns citizen physiological stress into real-time diagnostic infrastructure. Apple Watch sensors (HRV, heart rate, motion, noise, temperature) are processed on-device to generate privacy-preserving stress scores. These scores stream through a 6-agent orchestration chain (Ingestion → Mapping → Diagnosis → Simulation → Planner → Narrator) to identify heat islands, diagnose root causes (heat, noise, dangerous crossings), simulate interventions (shade canopies, parklets, longer crossing times), and rank solutions by biological impact per dollar. The result: a 3D digital twin of the city where red zones pulse in real-time, and city planners can click to see AI-generated intervention recommendations with predicted stress reduction.

How we built it

Edge Layer (On-Device Privacy): Apple Watch collects 6 signal categories (cardiovascular, gait, metabolism, environmental, thermal, GPS). A quantized stress classification model runs on-device (ZETIC Melange platform optimized for NPU deployment). Only three anonymized fields leave the device: ALS score (0.0–1.0), movement context, and H3 tile ID.

Multi-Agent Orchestration (Fetch.ai Agentverse): Built a 6-agent reasoning chain using Fetch.ai uAgents framework, registered on Agentverse with handle @superblock. Agents communicate via the Almanac discovery protocol and are accessible through ASI:One chat. The Ingestion Agent validates World ID proof-of-humanity, Mapping Agent aggregates into H3 hexagonal tiles, Diagnosis Agent infers failure modes, Simulation Agent runs what-if counterfactuals, Planner Agent ranks interventions by Biological Relief Coefficient, and Narrator Agent explains reasoning in plain language.

Backend (FastAPI + MongoDB Atlas): Python FastAPI service with 17,952+ packets persisted in MongoDB Atlas. Internal orchestration endpoint /agents/orchestrate/internal allows the skill agent to trigger the full reasoning chain without authentication for demo purposes.

Frontend (React + Deck.gl): Vite + React frontend with Zustand state management. 3D map visualization using Mapbox and Deck.gl H3HexagonLayer. Time slider replays stress patterns through the day. World ID IDKit v4 integration for proof-of-humanity verification.

Networking (Arista): WebSocket + REST fallback ensures low-latency data flow from edge devices to city intelligence backend. Multi-agent orchestration routes data through specialized agents.

Authentication (Auth0): Secures agent-to-user identity management for the backend API.

Public Exposure: Cloudflared tunnel exposes the local agent to Agentverse for ASI:One discoverability.

Challenges we ran into

Dynamic Endpoint Management: Cloudflared tunnel URLs change on every restart, requiring us to dynamically update the agent's endpoint configuration and Agentverse profile to ensure the agent remains discoverable.

SSL Certificate Configuration: Initial agent registration with Agentverse failed due to SSL verification errors, which we resolved by updating Python's certificate store and configuring proper SSL contexts.

Authentication Flow Design: Balancing security with agent accessibility required us to create separate internal endpoints for agent-to-backend communication while maintaining Auth0 protection for user-facing APIs.

Mobile Integration Planning: While we optimized the stress model for NPU deployment via ZETIC Melange, we identified that full on-device inference requires dedicated mobile app development—a longer-term technical path we documented for future implementation.

Accomplishments that we're proud of

Fully Functional Multi-Agent System: 6 agents registered on Fetch.ai Agentverse, discoverable via ASI:One chat with handle @superblock, responding to queries like "What's the best $10k fix for the heat island in DTLA?"

Privacy-First Architecture: Raw biometrics never leave the device—only anonymized stress scores and coarse location tiles are transmitted, with World ID proof-of-humanity ensuring sybil resistance.

Real-Time Digital Twin: 3D map with 2,791 tiles across Downtown LA, showing stress gradients from green zones (Grand Park) to critical red zones (Arts District, Skid Row).

18,000+ Packets in MongoDB Atlas: Durable telemetry persistence with 999 unique simulated users demonstrating scale.

World ID IDKit v4 Integration: Actual proof verification in frontend, not just mock integration—verified humans can submit data.

End-to-End Demo Workflow: From edge data ingestion through agent orchestration to ranked intervention recommendations, all working in a single cohesive system.

What we learned

Agentic Architecture Design: Multi-agent systems require careful protocol definition and state management. The separation of concerns across Ingestion, Mapping, Diagnosis, Simulation, Planning, and Narration agents created a modular, testable architecture that scales well.

Edge-Cloud Balance: Privacy preservation and real-time responsiveness require thoughtful architecture decisions. Keeping raw biometrics on-device while streaming aggregated scores enables both privacy and actionable insights.

Discovery vs. Registration: Agentverse offers two registration paths—API registration (free, sufficient for basic discovery) and Almanac contract registration (requires FET tokens, enables full ASI:One monetization). Understanding this distinction is crucial for production deployment planning.

Cross-Platform Integration: Connecting wearables, mobile devices, web frontends, and cloud backends requires careful API design and protocol selection. WebSocket for real-time streams, REST for control operations, and standardized data formats enable seamless interoperability.

What's next for Superblock

The "Clean Corridor" Expansion: We are integrating air quality (PM2.5) and humidity sensors to track the transition away from fossil-fuel-heavy transit corridors. This creates a data-backed mandate for expanding green transit and clean energy microgrids in the neighborhoods that need them most.

Direct-Action Micro-Payments: By implementing the Fetch.ai Payment Protocol, we will enable city planners—and eventually community DAOs—to fund hyper-local climate interventions (like cooling centers or solar-powered shade structures) directly through our agentic marketplace.

Almanac Contract Registration: Acquire FET tokens to register the agent on the Almanac contract for full ASI:One discovery and monetization capabilities.

Real User Testing: Deploy to a small neighborhood with actual Apple Watch users to validate thermal stress detection and intervention recommendations.

City Pilot Program: Partner with Los Angeles Department of Transportation and Planning Department to pilot in a real corridor, with actual shade canopy and parklet deployments based on AI recommendations.

Expanded Sensor Suite: Add air quality (PM2.5, NO2) and humidity sensors to improve diagnosis accuracy and detect compound stressors.

Multi-City Deployment: Extend to other cities facing extreme heat (Phoenix, Houston, Miami) with local calibration and neighborhood-specific interventions.

Built With

Share this project:

Updates