AeroFlux: Stage 2 — Real-Time Inference
AeroFlux: Fuse. Predict. Explain.
A live streaming inference and explanation layer that carries the Stage 1 model into production: FAA SWIM position feeds stream through Kafka into a continuously updated state store, active flights are scored as their upstream state changes, and results surface through a web interface backed by an LLM agent that explains why a delay is propagating — not just that it is.
Core technical components:
- SWIM → Kafka ingestion — FAA SWIM flight messages parsed and published to
swim.raw.flight(Kafka in KRaft mode), decoupling feed volatility from downstream consumers - Streaming fusion layer — raw messages normalized and joined against airport metadata and prior-leg aircraft state in Postgres, reconstructing the rotation features the Stage 1 model depends on in real time
- Live inference service — the trained XGBoost model scored against current flight state on a rolling cycle, producing continuously refreshed delay risk rather than one-shot predictions
- LangGraph explanation agent — retrieval-augmented reasoning over pgvector, turning a model score into an operational explanation of which upstream conditions are driving it
- Dash/Plotly interface — live flight map, risk surfacing, and agent interaction in a single view
- Containerized deployment —
aeroflux-ui,aeroflux-agent,aeroflux-agent-pgvector, andaeroflux-caddyunder Docker Compose on AWS Lightsail, fronted by Caddy with automatic TLS
From Batch to Stream
Stage 1 proved the signal exists; Stage 2 proves it can be delivered while the decision window is still open. The hard problem shifts from model accuracy to state freshness: the prior-leg feature that carries 51% of model importance is only useful if upstream state is current. The system is therefore built around continuous reconstruction of aircraft rotation state, not periodic retraining.
The Explanation Layer
A risk score alone doesn’t change a decision. The agent layer answers the operational question — what is driving this, and what is downstream of it? — grounding responses in retrieved flight and airport context rather than free-form generation. This is the piece that turns a prediction into an awareness layer.
Mini Demos
Walkthroughs of each part of AeroFlux, from the interface down to on-prem operations.
AeroFlux Dashboard View
The main application surface — navigation, layout, and how the pieces fit together.
Live Flight Map & Network Overview
Active flights streaming in from SWIM, rendered against the national airport network.
Live Inference
Delay risk scored continuously against current aircraft state, refreshing as upstream conditions change.
Model Performance Page
Stage 1 validation results surfaced in-app — metrics, feature importance, and holdout behavior.
Full Integration
End-to-end: ingestion through fusion, inference, and agent explanation in a single pass.
Cloud Monitoring & Observability
Container health, resource utilization, and pipeline metrics for the deployed system.
Operations On-Prem
Local stack operation — Kafka, Postgres, and pgvector running outside the cloud deployment.
Project Setup
Coming soon.
Architecture
Try It & Read More
Open the live system — real-time flight map, delay risk, and agent explanations
AeroFlux Stage 2 repo on GitHub — ingestion, fusion, inference, agent, and deployment configuration