FlightSense.
Flight prediction, from data to conversation. An end-to-end platform connecting machine learning, weather data, and a natural-language interface.
From raw data
to useful answers.
- 01 / DATAAmazon S3Historical records
- 02 / MODELSageMakerTraining & inference
- 03 / INTERFACEStreamlitDashboard on EC2
01 / The problem
A delay rarely travels alone.
An aircraft’s earlier delays can affect later flights in its rotation. FlightSense explores those propagation patterns alongside schedule and weather information, then makes predictions accessible through a dashboard and conversational assistant.
The research evaluates models using 7.07 million BTS 2018 On-Time Performance records. Its three-version framework studies schedule-based features, aircraft rotation-chain features, and NOAA weather features.
I am a co-author of the FlightSense preprint, alongside Aditi J. Shelke, Renuka J. Shelke, and Nitin P. Hazarani. The system description and results here are attributed to the project’s paper and repository.
02 / The system
The model is only one piece.
The repository documents a platform combining data storage, SageMaker training and inference, a Streamlit dashboard on EC2, weather ingestion, and an Amazon Bedrock Nova Micro assistant.
- 01Flight data
- 02Amazon S3
- 03SageMaker
- 04Dashboard
BTS flight data is stored in S3. XGBoost models are trained in SageMaker, and the prediction endpoint serves the Streamlit dashboard on EC2.
Read all three paths
Prediction. BTS flight data is stored in S3. XGBoost models are trained in SageMaker, and the prediction endpoint serves the Streamlit dashboard on EC2.
Weather ingestion. The repository describes weather ingestion through AWS Lambda into Amazon RDS. That data supports the platform's weather-aware experience.
AI assistant. The Streamlit interface uses Amazon Bedrock Nova Micro and data-query tools to answer natural-language questions with project data.
Architecture reconstructed from the repository README. Explore documented paths; this is an explanation, not a live prediction service.
03 / Engineering decisions
Better signals before bigger models.
Represent the aircraft’s journey
The paper’s feature-engineering framework models delay propagation through aircraft rotation chains rather than relying solely on schedule features. Tail-number tracking connects upstream flight history to later flights.
This adds useful operational context, but also increases the importance of which information is available at prediction time. A strong offline result is not enough without careful feature availability and serving design.
Make prediction data usable in conversation
The repository documents an Amazon Bedrock assistant with data-query tools. It connects natural-language questions to the platform’s data rather than presenting a language model as the source of flight facts.
This architecture adds flexibility to the interface while introducing additional dependencies: data freshness, query-tool behavior, and model responses all need to be handled deliberately.
Decisions summarized from the paper and repository. Tradeoff observations are engineering interpretation.
04 / Evaluation
Results with context.
Reported in the FlightSense paper after integrating NOAA weather features. Dataset: 7.07M BTS 2018 flight records.
ROC-AUC measures a classifier’s ranking performance across thresholds; it is not the same as classification accuracy. This is the paper’s reported evaluation, not an independently reproduced benchmark on this website.
The repository README describes version 1 (0.732 ROC-AUC) and version 2 (0.875). The paper reports version 3 (0.879). Those figures retain their version and source labels; this page does not assume every evaluation condition is interchangeable.
Source: FlightSense, arXiv:2605.07364, submitted 8 May 2026; revised 16 May 2026.
05 / Scope & limitations
Inspectable work.
Honest boundaries.
The paper describes model evaluation and an AWS deployment; the public repository contains notebooks, dashboard code, assistant code, and data tools. The repository’s weather integration and the paper’s NOAA feature evaluation are separate documented contexts.
This case study presents those artifacts and research results. It does not claim a currently available live endpoint, independently verified production performance, or sole ownership of the collaborative research.
Follow the work back to its source.