KEZEL
analyticsAug 2026

The questions don't churn

Insights Studio was introduced at Bengaluru Tech Summit 2025. The design decision behind it is older: organize analytics around the question, because the question is the only part that lasts.

Analytics tooling mostly organizes by artifact. Dashboards live here, notebooks there, the model registry somewhere else, and the user is expected to know which artifact their problem belongs to. Artifacts churn. Chart libraries change, model families go in and out of favor, last year’s notebook does not run.

The questions do not churn. A business asks the same three, quarter after quarter: how much, which ones, when. KEZEL Insights Studio ships one engine per question.

How much

The regression engine fits the relationship between a lever and an outcome, then lets you move the lever and watch the outcome respond before you commit to it. The simulations recompute against the boundary’s own tables, so the answer describes your data as it stands rather than an extract from whenever the model was last refreshed.

Which ones

The classification engine puts the riskiest items on top. Under the surface, 8 model candidates are evaluated, with and without outlier correction, and the scores that come out carry their reasons. A reviewer sees why row one is row one; reading the queue is itself a logged event.

When

The time-series engine, built on the Prophet and ARIMA model families, projects forward with a confidence band attached. The band is not decoration. A timing decision made off the center line alone is a guess wearing a chart; the band is what tells you how much the projection deserves to be trusted.

fig. 01 · actuals, then projection · the band is part of the answer · sample data

The fourth question

There is a question the engines leave alone: why. Attribution needs context no model in the boundary holds, and an engine that answers “why” with confidence is generating narrative, not analysis. That work stays with an analyst. The engines’ job is to hand the analyst a defensible how-much, which-ones and when to reason from.

All three engines run inside the analytics module, beside the system of record, with results read under policy. The analytics module page shows each one working on sample data, alongside the dashboards (VizEngine) and plain-language queries (LLMCore) that surround them.

written by the KEZEL team at DBTEZ · all articles · book a demo