TAE Analyst Agent

Built for Tractafric Equipment's finance and analytics team, who previously answered these questions by hand in Excel. The real challenge wasn't wiring an LLM to a database: it was three specific traps in the data that break naive text-to-SQL.

The schema has 50 ambiguous columns, including three different "major class" fields, four GPM fields, and percentages stored as fractions. Hand-written semantic SQL views and schema cards mean the model never touches the raw table directly.

Two Caterpillar programmes define support as a GPM% curve rather than a flat threshold. A naive join silently multiplies or mis-picks rows, misstating about 26% of total support (roughly $570K). A correlated-subquery band join fixes this, and it's specifically covered by the benchmark's "join" category.

Business French doesn't match the schema vocabulary: "pompes hydrauliques" has to resolve to PPC codes like HGP, HPM, and HRP across 4,653 French part descriptions. A brute-force NumPy vector index (no vector DB needed at around 5,500 vectors) plus a deterministic taxonomy graph in NetworkX handles this.

Everything runs fully offline: SQLite in read-only mode, a local Ollama model (qwen3.5:9b), no cloud calls, no data leaving the machine. A 100-question gold-set benchmark, with ablations removing the views, resolution, and graph one at a time, shows the semantic-view layer alone is worth about 47 accuracy points, and that the smallest model in the sweep is also the most accurate: reasoning mode and a 30B model both score worse than the 9B default with reasoning off.