Grounded incident orchestration

AEGIS

A supply-chain digital twin that lets an LLM coordinate tools without letting it invent the numbers.

Role: System design · Data engineering · Full stack
Year: 2026
Repository: Private

01 / Interactive system slice

Inspect the workflow

Public sandbox · real engine with clearly labeled cached and preview states

REPLAY

How this was decided

Use the steps to inspect the trace

Collect active threats

The engine grounds the replay in the curated Memphis hurricane scenario.

PREVIEW

LIVE NETWORK

Real project map surface · curated incident state · controls above adjust the inspection view.

02 / Engineering evidence

Claims with boundaries

Derived from the local implementation and offline artifacts.

Network

40+ nodesFacilities and logistics edges in the demonstration twin.

Tool surface

8 toolsTyped operations the orchestrator may invoke.

Evaluation

10 scenariosOffline incident scenarios in the checked-in evaluation suite.

AEGIS incident map from the project interface — Real project capture · incident-map state · source interface shown at its native aspect ratio.

Problem

Why this exists

Incident response is spatial, probabilistic, and constrained. A fluent narrative is not enough: each recommendation must be traceable to network state and solver output.

System

How it is bounded

The orchestrator requests spatial intersection, inventory risk, and rerouting calculations from deterministic tools. A provenance ledger binds every reported quantity to the call that produced it.

Outcome

What the demo proves

The Memphis replay below exposes the boundary: language sequences the work; geospatial queries, simulation, and optimization own the calculations.

Implementation surface

PythonFastAPIPostGISSciPyNext.js

Building reliable data products?

Let's talk through the system.

Start a conversation