Inspiration

Cities bring us physically closer, yet emotionally distant. We pass thousands of strangers every day, moving like isolated islands in an urban archipelago.

I was inspired by Coca‑Cola’s “Open Happiness” philosophy and the PERMA model from positive psychology. The core question was simple: What if AI could go beyond answering questions and actively orchestrate meaningful micro‑connections between strangers?

Mathematically, urban loneliness felt like:

[ \text{Isolation} = \frac{\text{Physical Proximity}}{\text{Meaningful Interactions}} \to \infty ]

Open Ground (共鸣场) was created to flip this equation — using agentic AI to transform everyday observations into collaborative micro‑missions that spark gentle social encounters.


What it does

Open Ground is an autonomous urban mission orchestrator that transforms everyday observations into shared missions that require two strangers to complete together.

Agent Flow:

  • Observe → User shares an urban detail (“There’s a faded bulletin board near the café.”)
  • Gather Context → Multi‑tool chain:

[ \text{Location} \to \text{Weather} \to \text{Time} \to \text{Nearby} ]

  • Generate Mission → Personalized micro‑tasks with verification criteria and secret codes
  • Verify Completion → Gemini Vision analyzes submitted photos
  • Self‑Correct → Mission adapts when conditions change (weather, time, environment)

The agent loop follows an iterative optimization pattern:

[ \text{Mission}_{n+1} = f(\text{Mission}_n, \text{Context}, \text{Feedback}) ]

The result: tiny, real‑world social sparks that make cities feel less like islands.


How we built it

Tech Stack

Layer Technology
AI Core Gemini 3 Flash (JSON Schema + Vision)
Frontend Vite + React + TypeScript
Tools Browser Geolocation, Open‑Meteo API, OpenStreetMap
Deploy Vercel

Architecture

This is a true agentic system, not a prompt wrapper. It demonstrates:

  • Multi‑step tool orchestration ((n \ge 4))
  • Verification loops with self‑correction
  • Real‑time thought streaming for transparency

Challenges we ran into

Environment Variables

Vite requires import.meta.env.VITE_* instead of process.env.*. Debugging the blank page consumed significant time.

Vercel SPA Routing

Client‑side routing conflicted with Vercel defaults and required custom vercel.json rewrite rules.

Agent State Management

Maintaining consistent context across multi‑step tool calls while adapting to real‑time conditions:

[ \text{State}t = g(\text{State}{t-1}, \text{ToolOutput}_t, \text{UserInput}) ]

Solo Development Under Time Pressure

Concept → prototype → deployment in 5 hours required rapid iteration and decisive trade‑offs.


Accomplishments that we're proud of

  • Built a fully agentic system — not a prompt wrapper — in under 5 hours.
  • Designed a multi‑step tool chain that adapts to real‑world conditions.
  • Implemented Vision‑based mission verification with self‑correction loops.
  • Created a transparent thought‑streaming UI that makes the agent feel alive.
  • Translated an emotional philosophy (urban resonance + micro‑connections) into a working product.
  • Proved that one person + modern AI tools can build meaningful social technology rapidly.

What we learned

  • True agentic AI = Tool Orchestration + Verification Loops + Self‑Correction
  • Gemini Vision is remarkably capable for real‑world verification
  • Transparent reasoning builds user trust
  • Modern AI tools enable rapid, meaningful product creation
  • Biggest insight:
    > An agent is not defined by what it knows, but by what it can do — and how it adapts when things go wrong.

What's next for Open Ground

  • Multi‑user collaborative missions
  • A “city resonance map” visualizing micro‑connections across neighborhoods
  • More contextual tools (events, transit, density, foot‑traffic)
  • Anonymous, serendipitous social graph
  • Long‑term agent memory for evolving city narratives
  • Deploying Open Ground in real public spaces as an urban social experiment

Built With

Share this project:

Updates