Sales To User Data Q&A Assistant Agent

Built on real Caterpillar-dealer aftermarket sales data from Tractafric Equipment the same dealer network behind my financial process automation project this project asks whether a sales manager can question the data directly instead of building a pivot table. A FastAPI backend turns a plain-language question into a single read-only SQL query, run through a local Ollama model (no cloud APIs) and checked by a two-tier guardrail layer a sqlglot-based safety gate plus a semantic linter before it ever touches the 3.1-million-row DuckDB warehouse. The real engineering problem wasn't prompting the LLM; it was the data itself: five verified traps that make the "obvious" SQL confidently wrong with no error, including a per-row running-total column that overstates revenue by 6.4x if summed naively, an 83%-empty reporting table where COUNT(*) is not a transaction count, three years of data silently truncated to exactly 1,000,000 rows and unevenly by country, and geography encoded only inside a French-language dealer-code string. Every answer streams back live with its exact SQL, a narrated explanation, and caveats automatically attached whenever the query touches unreliable data, plus one-click export to CSV, Markdown, HTML, Excel, or PDF, auto-generated charts, and a self-serve data-exploration tab. A 14-configuration evaluation matrix 6 local models times reasoning on/off, plus 2 baselines, including adversarial cases that must be refused outright benchmarks accuracy, latency, and correctness to recommend the best model for the hardware it runs on.