FEDERAL DEMOGRAPHIC REGISTRIESLIVENATIONAL LABOR DATALIVEPUBLIC DISASTER INDEXLIVECOMPETITOR INTELLIGENCELIVEREAL-TIME DEMAND SIGNALSLIVEESTABLISHMENT CATALOGUESLIVEGEOSPATIAL REFERENCE LAYERSLIVEHOUSING & INCOME DATASETSLIVEFEDERAL DEMOGRAPHIC REGISTRIESLIVENATIONAL LABOR DATALIVEPUBLIC DISASTER INDEXLIVECOMPETITOR INTELLIGENCELIVEREAL-TIME DEMAND SIGNALSLIVEESTABLISHMENT CATALOGUESLIVEGEOSPATIAL REFERENCE LAYERSLIVEHOUSING & INCOME DATASETSLIVEFEDERAL DEMOGRAPHIC REGISTRIESLIVENATIONAL LABOR DATALIVEPUBLIC DISASTER INDEXLIVECOMPETITOR INTELLIGENCELIVEREAL-TIME DEMAND SIGNALSLIVEESTABLISHMENT CATALOGUESLIVEGEOSPATIAL REFERENCE LAYERSLIVEHOUSING & INCOME DATASETSLIVE
← THE ARCHIVE
003ProductFebruary 20265 min

Inside the Advisor: four stages, one narrated brief.

The Advisor isn't a chatbot. It's a research pipeline with four explicit stages. Here's what each one does, and why we refused to skip any of them.

E
Engineering
Shipping the pipeline since v0.

Not a chatbot

When we first prototyped an "AI advisor," we did what everyone does: we strapped a model to our database and let it free-form responses. The answers sounded plausible. They were also frequently wrong — confident at the edges of what the data could actually support.

We killed the free-form prototype and replaced it with a four-stage pipeline.

Stage 01 — Classify

The model reads your natural-language brief and extracts three things: service type, geography (specific address, city, metro, state, national), and analytical scope. Nothing else happens until this step confirms itself.

Stage 02 — Discover

Based on the scope, we either skip discovery (if you gave us an address), run a lightweight candidate search (if you gave us a city), or run the full national opportunity pass. The output is a ranked candidate list — ZIPs that deserve drilling.

Stage 03 — Drill

Parallel fetch across every registered source in our pipeline: competitor footprint, wage and employment data, demographic tables, disaster exposure, establishment density, demand curves. We pillar-score each candidate against your service-type weights. This is where the actual research happens.

Stage 04 — Synthesize

Only now does the model generate narrative. It's asked to explain the top picks, name the tradeoffs, and flag anything anomalous — with each claim bound to a specific pillar score.

The reason for all four

You could get to an "answer" in one step. You'd also get lower accuracy, no source attribution, and no way to know where the model was improvising. The extra latency — usually 20 to 60 seconds — is the price of a brief you can actually take to a partner meeting.

— END · № 003
← BACK TO ARCHIVE
◈ THE DISPATCH

Like this issue?
Get the next one in your inbox.

Blog - Market Analysis Insights & Business Growth Tips | EllieReid