FDPIS: Flight Delay Propagation Intelligence System

Live demo

• Web app: https://fdpis.vercel.app
• Model API: https://fdpis-api.onrender.com (OpenAPI docs at /docs, health at /api/health)

The API runs on a free instance that sleeps after 15 minutes idle. The first visit wakes it, which takes about 45 s, and the site shows a "waking up" banner meanwhile. The demo serves the held-out test split (2025-07-19 … 07-31).

Research prototype. US BTS on-time data, Jan/Apr/Jul 2025. Not an operational system. Crew alerts use a hypothetical one-crew-per-aircraft model; notifications are simulated.

What it does

One late aircraft is never one late flight. FDPIS works in layers:

1. Data and rotation graph. Builds every aircraft's daily rotation from BTS on-time records.
2. Layer 3, primary risk. Ranks each departure by delay risk before the day begins (XGBoost, 32 features).
3. Layer 4, propagation. Pushes an observed delay down the rest of that aircraft's rotation using a hurdle gate plus quantile models, with an interval to depth 2.
4. Layer 5, actions. Checks the projected chain against FAA Part 117 or DGCA FDTL duty rules, kept as data with their regulatory sources, and raises dated crew alerts with simulated notifications.

Results (held-out test split)

All figures are from AUDIT_REPORT.md, which also covers the limitations.

Layer	Metric	Value
3	ROC AUC / PR-AUC	0.6924 / 0.3339
3	Precision at top 1% (carrier-weighted)	0.781 (0.696)
4	Gate AUC / cascade MAE	0.9272 / 8.35 min
5	Actionable recall at 85% precision (LIKELY only, FAA)	30%

Repository

Path	What
fdpis_*.ipynb	Data preparation, modelling, propagation, diagnostics and weather studies (notebooks 1–8)
api/	FastAPI model server and the Layer 5 rule engine (api/layer5/, rules in api/layer5/rules/*.json)
fdpis-web/	Next.js 14 frontend
deploy/	Deployment: slim serving build, Dockerfiles, parity and verification tools (deploy/README.md)
AUDIT_REPORT.md	Every reported number and how it was produced

data/ and results/ are produced by the notebooks and are not committed. The deployment's slim artifacts in deploy/backend/artifacts/ are committed.

Run locally

# API: full artifacts, after running the notebooks
uvicorn api.main:app --port 8000

# API: the deployed slim build (test split), from a clean clone, no notebooks needed
docker build -f deploy/backend/Dockerfile -t fdpis-api .
docker run --rm -p 8000:10000 fdpis-api

# Frontend
cd fdpis-web && npm ci && NEXT_PUBLIC_API_URL=http://localhost:8000 npm run dev

Team

Shrish Singh Sourya, Ashish Pathak, Ssunavya Jain. Guide: Prof. Anuradha G., VIT Vellore.