Operational leaders often receive several dashboards but still have to filter through them manually to determine what changed and where to investigate. I built this project to demonstrate how a governed analytics layer can support both a traditional scorecard and a plain-language diagnostic experience.
The application uses 697,754 real domestic flight records for July 2026 from the U.S. Department of Transportation's Bureau of Transportation Statistics. The raw data contains flight dates, airports, carriers, arrival and departure performance, cancellations, diversions, taxi times, and attributed delay causes.
I created a streaming Python pipeline that reads the compressed source without loading the entire dataset into memory. It applies consistent metric definitions, creates national and regional weekly aggregations, and exports compact structured JSON. Raw flight data remains outside the repository.
The HTML dashboard reads the governed JSON and presents ten operational KPIs. Users can change the region and week, compare performance with the prior week, examine carrier and airport hotspots, and review the delay mix.
The diagnostic copilot answers three practical questions:
- What changed?
- What drove performance?
- Where should we focus?
Every statement is calculated from the visible scorecard data, which makes the output traceable and reduces the risk of unsupported explanations.
- Automated data-quality checks and reconciliation totals
- A warehouse model with versioned metric definitions
- Role-based access and enterprise authentication
- An LLM layer with retrieval and tool calling for flexible follow-up questions
- Evaluation sets that test factual accuracy and explanation quality
- Monitoring for data freshness, pipeline failures, and model behavior
This project reflects how I approach applied AI: establish reliable data and metrics first, then build an interface that helps people reach decisions faster and understand the evidence behind the answer.