LateBird ✈️

A full-stack flight reliability tool that answers one question: is this flight usually late? No login required! Just enter route or a flight number and get on-time rates, delay causes, and the best hour to depart, all derived from the federal on-time record.

---

🔍 Features

• ✈️ Flight Reliability Lookup

  • On-time rate, delay causes, and the 20 most recent operations for any covered flight number.
  • Expand any operation for a plain-English reading of what happened.

• 🕐 Best Time to Fly

  • Ranks all 24 departure hours on a route and recommends a window.
  • LAX–SFO departing 06:00–07:59 is 86% on time against a route average of 47%.

• 🧠 ML-Based Delay Archetypes

  • KMeans clustering on normalized BTS data assigns each operation a labeled archetype.
  • An LLM phrases the finished fact set — it never sees raw numbers and cannot invent a statistic.

• 🛡️ Sample-Guarded Statistics

  • No rate is displayed below 10 completed flights. "100% on time" from 3 flights is a bug, not a feature.

• 🗃️ Cached BTS + OpenSky Data

  • Historical schedules and delays come from BTS; OpenSky supplies recent actual operations only.
  • OpenSky has no scheduled-departure field, so its rows are never labeled as delays.

---

📊 Coverage

Statistics cover 430,977 flights across the 30 busiest US airports and 13 carriers, from May–June 2026. A route is answerable only when both endpoints are in that set, so the route picker offers exactly those 30 airports and nothing else.

On time means arriving within 15 minutes of schedule — the FAA's definition.

---

🧱 Tech Stack

Backend

• Spring Boot 3 (Java 17) — REST API, OpenSky ingestion, nightly aggregate job.
• PostgreSQL — stores flights, airports, carriers, and the route reliability aggregate.

Frontend

• Next.js 15 (TypeScript, App Router) — Server Components by default, typed fetch wrappers.
• shadcn/ui + Tailwind CSS v4 — semantic delay palette, light and dark themes.

Machine Learning

• Python (Flask) — microservice that classifies one operation and phrases it.
• scikit-learn — StandardScaler + KMeans.
• Pandas — preprocessing and aggregation.

Ownership is strict: Python owns the model, Java owns the API and database. Spring talks to Flask over HTTP rather than reimplementing the model.

---

🧪 Machine Learning Pipeline

• BTS CSV → Postgres → StandardScaler + KMeans → labeled clusters + thresholds.json.
• Rule thresholds come from corpus percentiles, never hand-tuned numbers.
• The trained .pkl files are committed; the deploy target does not retrain.

---

📥 Getting Started (Dev)

# Database
docker compose up -d

# ML pipeline (one time, populates Postgres)
cd ml
pip install -r requirements.txt
python pipeline.py

# ML Microservice — port 5001, since macOS AirPlay owns 5000
cd flight-analyzer
pip install -r requirements.txt
flask --app app run --port 5001

# Backend (Spring Boot) — port 8081
cd server
./mvnw spring-boot:run

# Frontend — port 3000
cd frontend
npm install
npm run dev

---

About

Live demo: flight-reliability-analyzer.vercel.app — hosted on free tiers, so the first request after an idle period takes up to a minute while the API wakes; the interface shows a loading state rather than an error.

This project was created by Benjamin Yang (me) yangbenjamin19@gmail.com.

📄 Disclaimer

LateBird is not endorsed by, affiliated with, or representative of the US Bureau of Transportation Statistics, the Federal Aviation Administration, OpenSky Network, or any airline. Statistics are derived from publicly available federal on-time performance data and describe historical patterns only — they are not a prediction about any future flight.