The numbers refreshed. The explanation needed to refresh too.

I extended an existing dashboard with tested hypotheses, traceable calculations, and a separate AI narrator.

A closer look at the workExplore the case

Demand Gen AnalyticsSystem overview

Refresh inputs

Preserve source data and distinguish automatic from manual fields.

Calculate and test

Compute metrics and evaluate 33 explicit hypotheses.

Reconcile

Independently recompute key numbers and assess freshness and sample size.

Explain

Show the evidence trail; let AI narrate only the checked findings.

Process illustration · Read the build decisions below

01 / The problem

Fresh numbers need a current explanation

A dashboard can look current while its written conclusions still describe an older data pull. I found that metric refreshes and insight generation were separate processes.

The team needed findings it could inspect: which number changed, how it was calculated, what threshold mattered, and whether the data was complete enough to act on.

02 / The build

How I built it

I put calculations in deterministic code. The analytical layer computes funnel conversion, sales velocity, win rate, source effectiveness, and period comparisons.

A registry of 33 hypotheses evaluates the current data. Each returns supported, refuted, or inconclusive, with the evidence and a recommended action. A separate check independently recomputes key numbers.

The interface shows the source, formula, threshold, and confidence behind a finding. An optional language-model narrator explains the checked output without recalculating the numbers.

My role

I built the analytical extension on a colleague's existing dashboard. My work covered additional metrics, the hypothesis engine, reconciliation checks, evidence trails, and the new insights interface.

03 / In practice

A conclusion you can trace

Observation

Example

Qualified leads: 100 last period, 80 this period

Calculation

Example

(80 − 100) / 100 = −20%

Rule

Example

Flag a decline of at least 10%

Check

Example

Recompute the count; inspect freshness and sample size

Action

Example

Investigate the channels contributing to the decline

Illustrative numbers, not company performance. The example shows how one hypothesis becomes a finding.

04 / The result

33testable hypotheses

14reconciliation checks

0LLM tokens for the calculation core

I built and deployed the extension with 33 hypotheses and 14 passing calculation checks. It gave each finding a visible calculation and evidence trail, and preserved the last known good values when a data pull failed.

The original dashboard was a colleague’s work. My extension covered the analysis and checking layer; some source fields still required manual maintenance and the narrator described the checked data snapshot.

Keep exploring

Content Machine

I built a content system that connects search research, company knowledge, writing, editorial review and publishing operations.

Next case