How it works
It doesn’t answer the question. It investigates it.
01
Connect
Read-only sources over MCP, each with a data classification.
02
Profile
Discover datasets, profile every field, suggest what each one means, and check quality.
03
Interpret
Work out what the question means against the data that exists.
04
Hypothesise
Generate candidate explanations worth testing.
05
Plan
For each hypothesis, design the analysis that would support or undermine it.
06
Test
Run the statistics and guarded queries. Record every result as evidence.
07
Judge
Mark each hypothesis supported, weakly supported, rejected or inconclusive.
08
Follow up
Refine what survived into sharper hypotheses and test those too.
09
Synthesise
Turn the surviving findings, with their caveats, into a narrative.
10
Present
Choose the charts and order, link every slide to its evidence, and produce the deck.
The statistics engine
The language model plans. The engine does the maths.
Every number in a Stack analysis comes from a deterministic statistical tool, recorded as evidence with its inputs and outputs. Nothing is estimated by a language model.
Describe
- Descriptive statistics
- Data-quality checks
- Profiles of every field
Compare
- t-tests (Welch)
- ANOVA
- Chi-square
- Two-proportion tests
- Effect sizes
Relate
- Pearson, Spearman and Kendall correlation
- OLS and logistic regression
Over time
- Trend
- Seasonality
- Period-over-period change
- Cohorts
Explain change
- Mix-shift decomposition
- Contribution analysis
- Period comparison
Stay honest
- Z-score anomalies
- Bonferroni, Holm and Benjamini–Hochberg correction
Security
Built for the security review.
Profiles, not raw rows
The language model sees dataset profiles and statistical results. Your raw records never go into a prompt.
Read-only by design
Queries are limited to read-only statements, one at a time, with row limits. Stack can’t change or delete your data.
Classified and encrypted
Every source carries a classification: public, internal, confidential or restricted. Credentials are encrypted at rest.
Your team, your access
Admins connect sources, analysts run investigations, viewers read the results.
Limits
An investigation, not an oracle.
- Hypotheses can be wrong. Stack is built to test them, not to declare them true.
- Data quality shapes the answer. Stack reports missing fields and broken joins before the charts.
- Where the data can’t support a claim, the deck says so.
- It drafts, people decide. Every analysis is a draft for professional judgement.
What does your leadership keep asking that takes weeks to answer?
See Stack take an open question, test every explanation against your data, and turn what survives into an evidenced deck.