Economy Dashboard

How this briefing is made

The short version: public data in, a rules-only economic call, one tightly constrained language model for the writing, and a validator that re-checks every number before anything is published. Twice a day, every day.

Where the numbers come from

The briefing tracks 26 indicators: 7 core series that drive the headline read, 11 supporting series, and 8 market rows. The backbone is FRED, the Federal Reserve Bank of St. Louis data service, grafted with same-day feeds where FRED lags: Treasury.gov for the 2-year and 10-year yields, the BLS for CPI and jobs as the agency publishes them, the New York Fed for the effective fed funds rate, CBOE for VIX, FiscalData for the federal debt, and a spot feed for gold. Every value passes a plausibility gate before it is used, and a value that fails it is blanked rather than narrated.

Who makes the economic call

Not the AI. A hand-written, rules-only classifier assigns one of five macro regimes from fixed thresholds on the core series, with hysteresis so the label does not flap between updates. When the rules do not describe the data, the dashboard says so and shows a neutral no-call state instead of forcing a verdict. The language model never re-derives or overrides that call.

What the AI actually does

One constrained language model call writes the briefing paragraph: schema-pinned, low temperature, narrating the regime the rules already chose. A failover model stands behind the primary and has taken over in production. Then a deterministic validator re-checks the writing:

• every number is verified against the source data within a fixed tolerance
• every date reference is checked against the input's own reporting periods
• indicator references are checked against a whitelist
• the text is scanned for investment-advice phrasing, which is banned
• length caps are enforced

An output that fails any check gets one retry, then a deterministic template card takes its place. The reader never sees an error, and a failed data run keeps the previous briefing on screen rather than publishing a degraded one.

Cadence

The pipeline runs twice a day on a schedule aligned to U.S. data releases, morning and afternoon Eastern time. It is not a real-time feed, it does not forecast, and it does not give investment advice.

Engineering notes

• More than 400 automated tests run the entire pipeline offline with injected fake data sources
• Committed JSON Schema contracts sit between pipeline stages
• Two runtime dependencies, total
• The page you are reading is a hand-built installable PWA: no framework, no bundler, no accounts, offline-capable

Built and operated by Modeo, an enterprise AI consultancy. Read the build case study, or see the live briefing.