Use deterministic code for computable analysis
Workbook parsing, schema profiling, linting, and questions with exact answers are handled in code. An optional LLM client is used only when explanatory prose is required.
Research · Analytics tooling
Tableau Copilot
A research prototype that analyzes local Tableau workbooks and data to provide linting, evidence-backed Q&A, and new visualizations.
- Role
- Georgia Tech research project · Public prototype scope
- Technologies
- FastAPI · Tableau TWB · SQLAlchemy · Pandas · React
01
Context
An analyst may know the question but still need specialized tooling to translate it into workbook structure, fields, and visualization changes. This prototype reads local .twb and CSV/XLSX files without Tableau Server and keeps the data and calculations behind each automated result inspectable.
02
Engineering
An XML-based TWB parser reads workbook, dashboard, worksheet, and datasource relationships, while a data profiler summarizes the local schema. Nineteen lint rules and twelve question-bank queries record both results and calculations in an evidence pack. New visualizations are written to a separate patched TWB with derived-dataset lineage and worksheet metadata.
My contribution
- Contributed to the prototype implementation and evaluation while preserving workbook and dataset context.
- Helped connect automated linting, question-bank Q&A, and dashboard augmentation into one flow.
Key technical decisions
Return calculations with each result
The evidence pack records source paths, computed metrics, queries, and limitations instead of returning only a chart. Users can inspect both the proposed visualization and the intermediate calculation.
03
Validation
Bundled TWB files and local data fixtures verify parsing, profiling, question-bank answers, linting, and workbook augmentation without external services.
Next steps
- Extend round-trip validation to a workbook corpus containing complex calculated fields, multiple datasources, and dashboard actions.
- The analytical appropriateness of a generated visualization still requires user judgment. A future evaluation should let analysts review source evidence, the proposed chart, and the patched workbook together.