Inspiration

Orbital congestion has shifted from a rare-collision problem to a continuous operations problem. Close approaches in LEO now occur constantly, and large constellations execute thousands of avoidance maneuvers per year.

The real cost isn’t the one catastrophic crash, it’s the recurring maneuver tax: fuel burn, downtime, and operational overhead.

We were inspired to treat conjunction management not as passive monitoring, but as an agentic economic optimization problem.

What it does

AstraGuard is a multi-agent orbital risk engine that turns conjunctions into autonomous economic actions.

It:

  • Ingests TLE data
  • Propagates 72h forward with SGP4
  • Detects and refines conjunction events
  • Computes probabilistic risk ( P_c )
  • Adjusts expected loss using Earth-impact context
  • Runs an AI consultant agent (Claude primary, Gemini fallback)
  • Decides: IGNORE / INSURE / MANEUVER
  • Executes payments via Stripe
  • Logs ROI and maneuver-tax savings
  • Visualizes everything in a 3D Cesium mission-control UI

This directly targets the Agentic AI Track, demonstrates Best Use of Data, and integrates a live FinTech execution layer.

How we built it

We structured AstraGuard as a layered agent system:

Physics Agent

SGP4 propagation + voxel hashing to scale conjunction detection.

Risk Agent

Refined TCA + assumed-covariance collision probability.

( \mathbb{E}[\text{Loss}] = P_c \cdot V )

Adjusted with geospatial context:

$$ \mathbb{E}[\text{Loss}]_{\text{adj}} = P_c \cdot V \cdot (1 + 2s) $$

Consultant Agent (Claude / Gemini)

Claude acts as the primary decision consultant, receiving deterministic economics and returning structured JSON decisions. Gemini provides fallback + multimodal expansion capability.

Commerce Agent

Stripe integration enables autonomous premium binding and parametric conjunction insurance.

Value Agent

Tracks maneuver tax saved vs. baseline and emits value signals for ROI proof.

Every autonomy loop writes to a ledger — satisfying the hackathon requirement of proving real-world value.


Challenges we ran into

  • Avoiding ( O(N^2) ) pair explosion when screening thousands of objects
  • Ensuring accurate ECI -> ECEF transforms for Cesium visualization
  • Constraining LLM output to deterministic, auditable decisions
  • Moving from “cool 3D map” to measurable economic optimisation
  • Designing a baseline policy to prove maneuver reduction

Accomplishments that we're proud of

  • A working 72-hour screening engine at scale
  • A closed-loop autonomous decision system
  • Deterministic economic guardrails around LLM reasoning
  • Live Stripe commerce integration
  • An ROI ledger proving maneuver-tax savings
  • A mission-control UI that demonstrates agent actions in real time

We built not just a dashboard, but an autonomous orbital operations stack.


What we learned

*The bottleneck isn’t collision prediction, it’s decision throughput.

  • Conservative threshold policies create unnecessary maneuvers.
  • Agentic AI must be paired with deterministic guardrails to be trusted.
  • Judges reward systems that act and measure, not just analyze.

What's next for AstraGuard

  • Fleet-level rolling-horizon maneuver optimization
  • Multi-agent negotiation between satellites
  • Real CDM ingestion + uncertainty modeling
  • Expanded parametric insurance products
  • EO-enhanced debris classification using Gemini

AstraGuard’s long-term vision is to become the autonomous economic layer for space traffic management, minimizing maneuver tax while keeping orbital shells stable at scale.

Built With

  • api
  • celestrak
  • cesiumjs-backend:-fastapi
  • claude
  • elevenlabs
  • google-gemini-api
  • numpy
  • opentelemetry
  • otlp/phoenix
  • pydantic-database/storage:-sqlite-(local-artifact-and-tle-storage)-core-computation:-numpy
  • python
  • sgp4-apis-&-integrations:-anthropic-claude-api
  • solana
  • sql
  • stripe
  • tle
  • typescript-frontend:-react
  • uvicorn
  • vite
Share this project:

Updates