01. Abstract & The Performance Bottleneck
In airline operations, computing the cost of an isolated flight delay is straightforward; modeling the multi-day network cascade it triggers across aircraft, crews, and passengers is an expensive computational bottleneck. SkyRipple is a full-stack discrete-event delay propagation engine and agentic recovery simulator prototype.
Operating on 582,304 real BTS flight records, 190,711 crew pairings, and 2.41 million synthetic passenger itineraries, the core simulation engine models how a localized disruption (e.g., a 2-hour capacity drop at ORD) carrying a direct cost of $181,376 triggers a modeled network-wide cascade of $942,080 — a 5.19x propagation multiplier.
The Optimization
By isolating state mutation from graph traversal and aggressively applying functools.lru_cache onto passenger loaders in Python, itinerary generation dropped from 110 seconds to 8 seconds. The FastAPI simulation endpoint achieved a steady-state resolution time of 16-20 seconds for a full national cascade.
167MB SQLite DB
Immutable Dec 2025 Schedule
16-20s Resolution
Full national cascade
5.19x Multiplier
Delay propagation cost
02. Agentic Arbitration & The LLM Boundary
To autonomously evaluate recovery strategies from cascading disruptions, SkyRipple deploys a 5-role Operations Control Center (Aircraft, Crew, Passenger, Gate, Duty Manager). Greedy heuristic engines often cause downstream secondary delays; SkyRipple simulates branch reassignments under operational constraints.
Gemini Proposes
gemini-flash-lite operates purely as an intent routing layer in /api/qa_router.py. It translates natural language inputs ('delay United 1234 by 90 mins and ground N24993 for 4 hours') into a strictly validated Pydantic schema.
Code Disposes
The LLM never computes downstream cascading costs directly. The validated structured intent is handed off to deterministic Python code executing against the pre-computed schedule graph.
Duty Manager Agent
Intercepts candidate recovery swaps, runs isolated simulations on each branch, suppresses illegal crew duty swaps, and selects the optimal reassignment. Reduced modeled disruption recovery costs by $172,854 compared to greedy heuristic logic on the ORD benchmark scenario.
OCC Multi-Agent Arbitration
03. The Multi-Day Midnight Seam & Guardrails
Airline schedules do not reset at midnight; delays carry over as displaced aircraft, illegal crew duty limits, and broken passenger connections. Phase 8 of SkyRipple implements a continuous operational month engine.
Continuous Physics
The engine models tight-gap overflows. For example, tail N464UA carried a 335-minute unrecovered delay through a 395-minute overnight maintenance buffer.
Spillover-Aware Costing
Standard bounded calculations arbitrarily truncate reactionary costs at calendar midnight. Spillover-aware calculation captures real-world impacts by crediting downstream delay reductions that calendar-based boundaries ignore.
Honesty Guardrails
If the simulation discovers no net-beneficial swaps, the system suppresses celebratory UI states and plainly confirms doing nothing was optimal. Modeled savings reflect verifiable schedule state improvements, eliminating phantom recovery claims.
04. Deployment Architecture: Client-Side Demonstrator
Serverless Precomputation
Rather than running the heavy Python discrete-event engine on live web servers, multi-day benchmark scenarios were precomputed and exported into static JSON artifacts for an interactive Next.js web client.
Browser-Side Analytics
Multi-megabyte JSON outputs were partitioned by operational domain and loaded on demand. Dynamic filtering, delay-tree navigation, and visual analytics execute client-side in 0ms, delivering full demonstrator depth with zero backend server overhead.