WHYSE
AGENTIC ANALYTICS

Let the data talk.

Point at a metric. WHYSE searches the whole dimensional space and ranks what actually explains the move — no hypothesis, no SQL, no waiting on the queue.

>
Revenue1,000800−200
West
−200
paid-social
−180
organic
−20
East
0
IF region = West AND channel = paid-social → 90% of the drop
SecondsTO INSIGHT
AggregatesONLY TO THE LLM
CommodityHARDWARE
THREE MINUTES

The whole argument, in three minutes.

Why dashboards stop at what, why SQL cannot get you to why, and what the engine does instead.

2:59

NO SQL · NO HYPOTHESIS · REAL PRODUCT FOOTAGE

THE PROBLEM

Dashboards tell you what happened. Nobody tells you why.

And when someone finally does, the room argues about whether the number is right.

Curiosity has a price tag

Curiosity has a price tag

Every new question is another full scan and more analyst hours. The more you want to know, the more it costs to ask.

Dashboards show the move, not the cause

Dashboards show the move, not the cause

A chart tells you revenue fell. It will not tell you which slice took it down.

Nobody trusts the answer

Nobody trusts the answer

Three analysts produce three numbers and the percentages add up to 107%. That is data paranoia, and more analysts make it worse, not better.

SQL is a retrieval language. It was never going to discover anything.

DASHBOARDS & SQL

Answer the question you already thought to ask.

WHYSE

Searches every segment and finds the one you didn't.

HOW IT WORKS

The last frontier of data science.

From raw tables to ranked answers.

01

Build the cube once

One pre-aggregation pass turns your warehouse into a cube of mergeable sketches. Every segment's stats become a lookup, not a scan.

1 PASS · EVERY SEGMENT
02

Ask in plain English

The model reads your schema, not your rows, and turns your sentence into a structured request. That request is shown back to you before anything runs.

> why did revenue drop last quarter?
measurerevenue
window2026-Q2 vs Q1
filtersnone
03

Get a ranked, traceable answer

A deterministic engine ranks every segment by how much it explains. Contributions sum to 100%, and the result is an if-then rule you can take into the room.

IF region = West AND channel = paid-social → 90% of the drop
IF region = West AND channel = paid-social → 90% of the drop

Same question, same data, same answer — every time.

CAPABILITIES

Six questions the engine answers without a query.

What moved the number

Ranks which segments drive a KPI and its change, in order of how much they explain.

CONTRIBUTION DECOMPOSITION

Where it's breaking

Finds the slices performing below their own baseline, not just below average.

ANOMALY & RARE EVENTS

What's missing or over-supplied

Separates a shortfall from a surplus and attributes each to the segments responsible.

DEFICIT & SURPLUS

What travels together

Surfaces feature combinations that co-occur — and the ones conspicuously missing.

PATTERN DISCOVERY

What changed shape

Flags segments whose distribution shifted, even when the average held steady.

TEMPORAL DRIFT

What predicts the outcome

Learns which attributes predict churn, default or conversion, as rules you can read.

PREDICTIVE SEGMENTATION

Drill-down, slicing and filtering are built in — same engine, no extra setup.

PROOF

The answer arrives while the decision is still open.

ANOMALY

The engine found the COVID cliff in NYC taxi data without being told to look

Observed 1,428 against a forecast of 2,114 — well outside the confidence interval, on a day nobody had flagged.

Read the investigation
Daily trip volume, 14 Mar 20202,114 forecast1,428 observed−686
Mid-fare, non-tipped
−686
All other segments
0
Mid-fare, non-tipped trips fell 32% below forecast on 2020-03-14 — outside the confidence interval
Under 7 min
TIME TO FIRST ANSWER
10k+
SEGMENTS PER QUESTION
0
QUERIES WRITTEN

Profiled on 100M+ records; formal benchmark in progress. Other figures are representative of internal runs, not customer results.

Two more investigations, including a fully public dataset: EIA Crude Production · DriveU

INTEGRATION

Three ways in. Zero rewrites.

MCP server

Your agent calls each capability as a named tool. We host; no infrastructure on your side.

Hosted service

We run the engine, you call a REST API. Fastest path to production.

Embedded library

Install as a dependency. Runs on your infrastructure; your data never leaves it.

Every answer ships as structured, explainable evidence your own agents can ground on — not just prose.

TRUST

Auditable by design.

Aggregates, not records

The engine computes over pre-aggregated cube cells — ranked segments and schema reach the model, not raw fact rows.

Deterministic results

The same question returns the same answer, and the path to it is auditable.

Your infrastructure

Runs on commodity hardware. Embedded deployment keeps everything in your environment.

FAQ

The questions your data scientist will ask.

Still have questions?

Talk to the team that built the engine.

Contact us

No, and nothing at this scale does. Pre-aggregation makes each lookup cheap, then the engine prunes: infrequent branches are dropped, and the search stops early once the ranking is stable. It's exhaustive within the cube's dimensions, not across all of mathematics.

Correlation, ranked well. The engine finds the segments that account for a change in observational data. Proving cause needs an experiment. We'll show you the slice to test — we won't claim the test is unnecessary.

Scanning thousands of segments guarantees some look significant by chance. Results are corrected for multiple comparisons and filtered by effect size, so a tiny segment with a dramatic-looking swing doesn't top the list.

Then it's a scan, and it's slow. The cube's speed comes from choosing dimensions in advance. Adding one means rebuilding, which is a background job, not a wait.

Yes, and it separates the two causes: the rate changing within segments, and the mix of segments changing underneath it. Conflating those is the most common way conversion analysis goes wrong.

It turns your sentence into a structured request — measure, window, filters. That's the one non-deterministic step, so the parsed request is shown back to you before the analysis runs.

The first cube usually builds within a day of connecting a source. Refreshes are incremental.

No, and the difference is the whole product. Text-to-SQL translates your question into a query — it still needs you to know which question to ask. WHYSE starts from the metric and searches the dimensional space to find which segments explain it. Different class of algorithm, not a better prompt.

You can, and it often works. Two catches. Generated code is not reliably correct, so somebody has to check it — which puts the work back on the analyst you were trying to free up. And it still only answers the question you asked; it will not hand you the segment you did not think to check.

No. It's automatic by default — point it at the data and it surfaces the drivers with zero setup — but analysts can steer the search and add domain knowledge, with the same numeric accountability. It fits their workflow instead of overriding their judgement.
GET STARTED

Ask your first why.

Bring one question you've been putting off. We'll build the cube and answer it on the call.