LB-05

Data Analyst

Answers business questions in plain language by writing, checking and running SQL on the company’s sales data.

Phase 1Flask · sync · Size M

What a visitor does Asks a sales question, then tries to drop a table

Problem

Owners wait days for simple numbers, and letting a model run SQL against a database is a real security risk.

Try it live

Ask which coffees lost the most repeat buyers last quarter. You get the SQL, the table, a chart and a plain explanation. Then try to make it delete data and watch each safety layer stop it.

Proves

Safe text-to-SQL, a semantic layer, and a measured choice of sync over async.

Tool use
Structured output
Text-to-SQL

Signal chain

resolve metrics
write SQL
parse + allowlist
EXPLAIN
run read-only
self-correct
chart + explain

Stack

Flask (sync) · gunicorn threaded workers · DuckDB · sqlglot · Vega-Lite · Vue

Engineering highlights

Six layers of SQL safety: a SELECT-only parse, an allowlist of tables, columns and functions, a plan check before anything runs, a locked read-only connection, a forced row limit and a query timeout.
Sync on purpose. Queries are short and CPU-bound, threads cover the model wait, and the code stays simple. The datasheet publishes a load test against an async build.
Execution accuracy measured on 100 question and answer pairs.

Table 1. Operating limits
Operating limit	Value
Questions per visitor per day	25
Query timeout	5 s
Row cap	1,000
Model calls per question (est.)	2–4

Status The evaluation board for this part is open. It starts on curated samples, and your own text runs live within your daily allowance. Open the evaluation board