Data Detective

README excerpt

# Snowflake hackathon

## Data Detective

Data Detective is an AI-powered data quality investigator built entirely inside Snowflake, with Claude as the investigator. It profiles weather observations from the Snowflake Marketplace dataset `SNOWFLAKE_CURATED_CORE_WEATHER_DATA`, flags anomalies, and explains why they happened. It also checks the data behind a wildfire risk score (`wfr-v4a`) to show whether a score can be trusted.

### What it does

1. **Profile and check.** Daily checks compare recent observations against historical baselines (per entity, metric and month) and flag null values, duplicates, stale records, volume drops, outliers and flatlined sensors. Wildfire score checks catch score jumps, high scores built on degraded inputs, and isolated alerts.
2. **Investigate.** For each finding, Claude runs a hypothesis-driven loop. It states a hypothesis, picks one read-only analysis to test it, reads the result, and updates its beliefs. The analyses cover entity profiles, time series, cross-metric comparisons, neighbor comparisons, network-level reporting and score breakdowns. It will also explicitly test unit and scale errors (such as F to C or mph to km/h) before it blames a sensor.
3. **Report.** Each investigation produces a structured, evidence-backed report: verdict (`DATA_ERROR`, `REAL_EVENT` or `INCONCLUSIVE`), severity, affected entities, dates and rows, what was ruled out, prioritized recommended actions, and a SQL check to detect recurrence.
4. **Review.** A S