DeadZones | Devpost

Inspiration

Everyone has experienced it — you're mid-navigation, mid-podcast, mid-search, and the bars drop to zero. By the time you notice, it's already too late. Current systems react after signal drops. We asked: what if an agent could see it coming?

What it does

DeadZone Agent predicts cellular dead zones along your route before you hit them, then autonomously builds a cited offline content pack — weather, road conditions, points of interest, local news — and delivers it to your phone while you still have signal.

When a second driver hits the same dead zone, their agent finds the cached pack and buys it from the first driver's agent instead of rebuilding it — a live agent-to-agent economy.

The demo flow:

Enter your route and departure time
Agent 1 predicts dead zones using real cellular signal data (CoverageMap) and route waypoints (Google Maps)
Agent 2 searches the web (Nimble), publishes a cited pack (Senso.ai), and caches it (ClickHouse)
Pack is delivered before signal drops
User B hits the same zone — their agent buys User A's pack in ~1 second for $0.02

How we built it

Two-agent architecture using raw OpenAI function-calling loop — no frameworks.

Agent 1 (Prediction): Google Maps Directions API + polyline decoding → CoverageMap batch API (real RSRP dBm per waypoint) → identifies dead zones below -105 dBm threshold
Agent 2 (Orchestrator): Checks ClickHouse cache → calls Nimble 4x in parallel (weather, roads, POIs, news) → publishes to Senso.ai → delivers via WebSocket
Payments: Coinbase CDP wallet on Base Sepolia + x402 agent-to-agent settlement
Frontend: Next.js + react-leaflet map with live agent reasoning stream
Observability: Datadog LLM Observability traces every tool call

Challenges

CoverageMap auth format wasn't documented — figured it out through trial and error
Wiring two independently-built agents with a clean handoff contract mid-hackathon
Switching from OpenAI to OpenRouter mid-build when API keys weren't available
Railway deployment with a monorepo containing Python backend + Next.js frontend

Accomplishments

Real signal data from CoverageMap on a live route — not simulated
Full agent-to-agent pipeline running end-to-end in under 10 seconds
Clean two-service architecture that any teammate could pick up and extend

What we learned

Context engineering isn't about prompting — it's about deciding what context to build before the moment of need. Our agent reasons about what a driver will need in 4 minutes, fetches it from the real web, and delivers it before the window closes.