Nebius × NVIDIA Hackathon 2026 Submission
Category: Agentic AI Systems • Infrastructure Resilience
Dispatch Sentinel is a real-time disaster-response digital twin that uses NVIDIA Nemotron 3.5 Lightning (via Nebius Token Factory) to triage unverified civilian emergency reports, applies Expected Value of Sample Information (EVSI) decision theory to determine when autonomous drone reconnaissance is justified, and re-optimizes convoy delivery routes using Google OR-Tools — all before a human watch commander authorizes any closure.
Flagship result (Scenario B — Critical Bottleneck Bridge):
| Policy | Planned Fulfillment | Sectors Served |
|---|---|---|
| Conservative (no verification) | 21.4% | 2/8 |
| Dispatch Sentinel | 100.0% | 8/8 |
| Advantage | +78.6 pp | +6 facilities |
This result is for a specific frozen disaster configuration. See
BENCHMARK_V3_RESULTS.jsonfor multi-scenario results.
Ham Radio / Social Media Reports
↓
[Nemotron 3.5 Lightning] ← Nebius Token Factory (NVIDIA NIM)
Structured triage + segment extraction
↓
[EVSI Governor]
Restricted-policy decision: scout or act?
EVSI_R = E_y[E_{x|y}[U(π_R(y),x)]] - E_x[U(π_0,x)] - C_recon
↓
[Drone Reconnaissance] (simulated, sensor η=0.999)
↓
[Human Operator Gate] ← required before any road closure
↓
[Google OR-Tools VRP] ← route reoptimization
↓
[Three.js 3D Digital Twin] ← live visualization
- Python 3.12+
- Docker (optional)
- Nebius API key (for live Nemotron inference)
git clone https://github.com/AmanM006/dispatch
cd dispatch
# Copy and fill in your Nebius key
cp .env.example .env
# Edit .env: NEBIUS_API_KEY=your_key_here
pip install -r requirements.txt
python -m uvicorn dispatch.twin_server:app --host 0.0.0.0 --port 8088
# Open http://localhost:8088docker build -t dispatch-sentinel .
docker run -e NEBIUS_API_KEY=your_key_here -p 8088:8080 dispatch-sentinel
# Open http://localhost:8088export PYTHONPATH=.
pytest tests/test_twin_api.py -v # 4/4 API unit tests
python tests/smoke_test_e2e.py # 20/20 live end-to-end test (calls Nemotron)Unstructured civilian reports (ham radio, social media) → structured {segment, confidence, is_rumor} JSON via Nemotron 3.5 Lightning.
Before dispatching a drone, the governor computes restricted-policy EVSI:
EVSI_R = E_y[E_{x|y}[U(π_R(y), x)]] - E_x[U(π_0, x)] - C_recon
If EVSI ≤ 0, scouting is not justified. The drone is dispatched only when information gain exceeds cost.
No road is ever closed without an operator reviewing the drone's sensor observation and explicitly authorizing. All actions are audit-logged.
One-click toggle shows what would have happened under the conservative "do nothing" baseline — same scenario, independently computed routes.
Real-time animated map showing convoy routes, destroyed bridges, isolated facilities, hospital markers, and live KPI telemetry.
| Endpoint | Method | Description |
|---|---|---|
/ |
GET | Serve twin_ui.html |
/api/state |
GET | Full simulation state |
/api/switch_scenario/{A|B} |
POST | Load disaster scenario |
/api/ingest_next_incident |
POST | Triage next report via Nemotron |
/api/dispatch_drone_recon |
POST | Dispatch reconnaissance drone |
/api/operator_action |
POST | Human authorization gate |
/api/counterfactual_state |
GET | Conservative baseline comparison |
/api/evsi_inspector/{id} |
GET | Full EVSI mathematical breakdown |
| Variable | Required | Description |
|---|---|---|
NEBIUS_API_KEY |
Yes | Nebius Token Factory API key |
PORT |
No | Server port (default: 8080 Docker / 8088 local) |
Never commit .env to version control. See .gitignore.
See BENCHMARK_V3_RESULTS.json for full multi-scenario results including:
- Scenario A (baseline urban routing)
- Scenario B (critical bottleneck — flagship scenario)
- Centrality and severity baseline comparisons
dispatch/
├── twin_server.py # FastAPI server + session management
├── twin_ui.html # Three.js digital twin UI (single-file)
├── controller.py # Google OR-Tools VRP route optimizer
├── triage.py # Nemotron integration + EVSI governor
└── scenario_loader.py # Disaster scenario definitions
tests/
├── test_twin_api.py # Unit tests (4 tests, no external calls)
└── smoke_test_e2e.py # End-to-end test (calls live Nemotron)
Apache 2.0 — see LICENSE.
Built for the Nebius × NVIDIA Hackathon 2026.