Project Overview

BlackBox is a data-analysis copilot designed around deterministic SQL execution and explicit query visibility. Users upload CSV or Excel files, which are ingested into an in-process DuckDB instance. A multi-stage LLM pipeline converts natural-language questions into a validated query plan, generates parameterized SQL, executes it in a read-only sandbox, and narrates the results. Privacy is enforced in code: the planner receives only schema metadata and up to five sample values per column; the narrator receives at most 20 aggregated rows. Features include automatic chart selection, contribution analysis, AI-proposed dashboards, and an offline mock provider for demos. The project ships with 495 backend tests, with end-to-end coverage comparing LLM answers against hand-computed Pandas results on fixture data.