Data analysis / Collaborative course project
NFL Data Analysis Agent
Natural-language questions, SQL-backed analysis, and a review step before generated code runs.
Analysis workflow and contribution
Developed with Arian Zarazua, this agent connects a natural-language interface with multi-season NFL datasets. It routes requests to SQL-backed tools or generated Python, retrieves supporting context, and presents outputs through a Streamlit interface.
The project brings together data ingestion, query generation, retrieval-augmented generation, and a human-in-the-loop workflow. Its engineering focus is the path from an informal question to an inspectable table, plot, or analytical result.
The written report documents the architecture and its limitations. The local SQLite database is rebuilt during initialization; the project is a course implementation rather than a production multi-user service.
How the system connects
Structure the records
Load the supplied datasets into SQLite and use schema context to support cross-table analysis.
Choose an analytical path
Combine LLM routing, registered Python tools, SQL, and FAISS-backed retrieval to connect a prompt with a method.
Inspect before execution
Use human approval and visible outputs to review generated code and analytical assumptions.
Generated outputs
Examples retained from the supplied project materials. These show the reporting workflow; they do not establish causal relationships or independently validated NFL findings.
Where review matters
The report describes a recurring challenge: code can run successfully while the analysis uses the wrong aggregation or an unsuitable set of features. Retrieval can also introduce irrelevant context, and SQLite limits concurrent use.
Those limitations make the approval and inspection steps part of the system design, rather than a claim that automated output is inherently trustworthy.
Report & supporting materials
The final course report, presentation, and development test logs.
Loading document…
