Six questions for an AI analyst
I put the assistant through six questions on demo data. A number it must prove, comparisons it must date, a headline it must catch, a query no widget can express, and a request it must refuse.
I build data platforms and AI systems where every number can be checked. By day I lead analytics and AI for a 15-location med spa group; this site is where I write about making that trustworthy: what the AI may build, when it must refuse, and how its work is checked against real numbers. The projects and the resume live here too.
plain language validated spec deterministic engine real numbers
I put the assistant through six questions on demo data. A number it must prove, comparisons it must date, a headline it must catch, a query no widget can express, and a request it must refuse.
Eleven plain sentences on demo data became an executive dashboard. The walkthrough, the specifications behind each widget, and the requests it narrowed or refused.
A dashboard number is the last inch of a long pipeline. What actually happens between an operational system's API and a figure an operator can check, told from production.
An in-product assistant that drafts dashboard widgets and answers analytical questions from governed data models. It never computes a figure; a deterministic engine does, and questions it cannot answer safely get a clear refusal instead of an improvised number.
AI-built widgets parity-tested against hand-built ones, to the cent
Point of sale, accounting, payroll, ad platforms and web traffic landed into a single dbt-modeled warehouse, fed by nightly API pulls and a live webhook feed with deletion reconciliation.
about 1.3 million invoice lines; ingestion reconciled to 99.87% against source
Tiered, rule-driven commission computation at invoice-line grain, reconciled line by line against the amounts actually paid. The reconciliation surfaced six-figure discrepancies traced to individual rate-configuration defects.
every amount traces to the invoice line and the rule that produced it