← ALL WORK · CASE 10
RAND STATISTICS
State-level statistics from many sources and formats, made queryable — pipelines that clean, models that summarise, and numbers that arrive with their receipts.
DATA PIPELINE — EVERY DATASET WALKS THIS PATH
ANOMALY FLAGGED → HUMAN REVIEW → EXPLAINED OR EXCLUDED — NO SILENT NUMBERS
01 — WARP
What had to be true first
RAND State Statistics arrived as a data problem wearing an AI costume: economic, healthcare and education statistics scattered across sources and formats, analysed by hand, slowly. The warp was ingestion — pipelines that pull, clean and normalise structured and unstructured state data into something a model can be trusted with.
02 — WEFT
What we wove through it
Onto clean data we wove the intelligence: LLM-generated interpretive summaries so researchers read findings instead of spreadsheets, anomaly detection that flags the outliers worth a human's time, and real-time querying tuned so an answer arrives while the question still matters.
03 — BEAT
How it was pressed tight
Public data carries governance obligations, so explainability was reviewed as a feature, not a virtue: summaries cite the data they draw from, flagged anomalies show their reasoning, and anything the system cannot ground gets surfaced for review rather than smoothed over.
04 — WEAR
Life in production
The platform serves researchers, policy-makers and public institutions in production, with manual analysis time down more than 60%. The models have been swapped as better ones arrived; the pipelines — the part that actually earned the number — haven't needed to change.
NEXT CASE — 11
WINSTON →