Skip to main content

AI data analyst

Upload a CSV, paste your schema, or describe the question and get analysis you can act on: a profile of a new dataset, the SQL to answer a question, a statistical test with the result read plainly, a dashboard spec, a narrative for the execs. Zeplik works from the data you give it, not made-up numbers, and it is honest about the line: it writes the query and the analysis, and it tells you which assumptions and figures to check before you report them.

Free to start. Nothing sends on its own.

Who it is for

How it works

  1. Share the data or the schema

    Upload a CSV, paste your table schema, or describe the dataset. Nothing is assumed. The analysis is built only from what you share, and the more real structure you give, the more precise the query and the result.

  2. Ask your question in plain English

    Say it plainly, for example "what is driving this drop" or "write the SQL for monthly revenue by region." Zeplik routes to the right data skill without you naming it, and you can steer the dialect, the method, and the output.

  3. Review the method, then report

    You get the query, the analysis, or the dashboard with the reasoning and assumptions shown. Check the sample, the method, and each figure against the source, then report it with confidence, all in one conversation.

What you can ask

Real asks in plain language, and the concrete deliverable you get back. Open the assistant to try your own; nothing runs until you send it.

Profile this dataset and tell me what stands out.

A first-pass profile of the CSV you upload: column types, null and outlier counts, distributions, and the relationships and anomalies worth a closer look.

Write the SQL to answer the question I will describe against my schema.

A correct query for your dialect, whether Snowflake, BigQuery, or Postgres, with window functions and date logic handled, and an explanation of why it is written that way.

Is this difference statistically significant?

The right test for your data with the hypothesis stated, the result read in plain language, the caveats on sample and assumptions, and what you can and cannot conclude.

Design a KPI dashboard for the metrics I care about.

A dashboard spec: which KPIs to show and how to define them, the layout and card design, and the charts to use, ready to build as an interactive HTML dashboard.

Turn these results into a story for the exec team.

A persuasive narrative over your analysis: the headline, the supporting points, and the chart choices, framed for the audience and the decision it drives.

Sanity-check this analysis before I share it.

A QA pass over your query and method: logic errors, wrong joins, silent filters, and unsupported conclusions, so the number holds up before it goes out.

What it can do

The assistant routes to the right skill for you. Each one below is a focused, ready-to-run workflow; follow a link to see exactly what it does.

Explore and understand

Get to know a dataset and catch problems before they bite.

  • Analyze dataAnswer a question of your data: counts, drivers, and segment comparisons.
  • Profile a datasetUnderstand a new table or CSV: types, nulls, and what is in the file.
  • Data contextCapture entity definitions, metric formulas, filters, and gotchas.
  • Analysis workflowChain explore, analyze, test, validate, story, and dashboard.
  • QA the analysisSanity-check a query, a method, and the conclusions before sharing.

Query and engineer

Write the SQL and move the data, from one table to billions of rows.

  • Fix and tune SQLWrite or fix SQL where dialect, window functions, and speed matter.
  • SQL from EnglishTurn a plain-English question into a correct SQL query.
  • Big-data framesPolars and Vaex for billion-row datasets and out-of-core work.
  • ClickHouseHigh-performance analytics: query optimization and schema design.
  • Data pipelinesScalable pipelines and warehouses with Spark, dbt, and Airflow.

Analyze and model

Bring statistical rigor and machine learning to the question.

  • Statistical rigorHypothesis tests, correlations, and outlier and anomaly detection.
  • Classical MLscikit-learn, statsmodels, Bayesian modeling, and explainability.
  • ML engineeringAirflow DAGs, dbt models, data-quality checks, and MLOps.
  • Data sciencePredictive models, experiment analysis, and business intelligence.

Visualize and report

Turn results into charts, dashboards, and a story that lands.

  • Build a dashboardAn interactive HTML dashboard with KPI cards, charts, and filters.
  • KPI designDecide which KPIs to show, how to define them, and the layout.
  • Python chartsMatplotlib, Seaborn, and Plotly plotting and interactive charts.
  • Chart from dataThe right chart from query results, a DataFrame, or a CSV.
  • Data storytellingTurn analysis into a persuasive narrative for a decision.

What a general chatbot gets wrong on data, and what is different here

A raw chatbot will confidently produce a number, a chart, and a conclusion with no data behind any of it. Zeplik works from the dataset and schema you provide, shows the query and the method, and is honest that a figure is only as good as its source.

The failure. Asked for a metric, a general model states a specific number as if it computed it, when it made it up.

How Zeplik grounds it. Zeplik works from the data you upload or the schema you paste, writes the query it would run, and shows its method, so a number is one you can reproduce, not one it invented.

The failure. A general model writes SQL for the wrong dialect, breaks on date functions, or silently drops rows with a bad join.

How Zeplik grounds it. It handles the dialect you name, gets window and date logic right, and explains the joins and filters, so you can see exactly what the query counts before you trust it.

The failure. A general model calls a difference significant with no test, no sample size, and no caveats.

How Zeplik grounds it. The statistics skills pick the right test, state the hypothesis and assumptions, and read the result plainly, including what you cannot conclude, so the claim is defensible.

What it does not do

Knowing the boundary is part of using it responsibly. This is where a person, not the assistant, has to own the call.

This is data analysis and draft code for your review, not a verified result. Every figure depends on the data you provide and the assumptions behind it, so check the query, the sample, and the method before you rely on a number, and confirm any statistic against the source before you report or act on it.

Frequently asked questions

Is the AI data analyst free?
Zeplik is free to start and every new account includes credits. Data skills run like any other Zeplik chat, so you spend only normal usage credits, with the cost shown before you send.
Can it query my database directly?
Not yet. It works from data you upload or a schema you paste, and it writes the SQL it would run so you can execute it yourself. It does not connect to your warehouse and run the query live.
Which SQL dialects does it handle?
It writes for the dialect you name, including Snowflake, BigQuery, and Postgres, and handles the date functions, window functions, and quirks that differ between them.
What do I need to give it?
For analysis, upload the CSV or dataset. For SQL, paste your table schema and the question. For statistics, share the data and what you are testing. For a dashboard, list the metrics you care about.
Can it do real statistics, not just averages?
Yes. It runs hypothesis tests, correlations, anomaly detection, and classical machine learning with scikit-learn and statsmodels, and it states the assumptions and caveats so the result is defensible.
Can it build a dashboard?
Yes. It helps you choose and define the KPIs, then produces an interactive HTML dashboard with cards, charts, and filters from the metrics and data you provide.

How to use it

  1. Upload the CSV, or paste the table schema if the data lives in a warehouse.
  2. Name the SQL dialect and any business rules, such as how you count an active user.
  3. Ask the question in plain words, for example "is the new checkout converting better?"
  4. Read the method and the assumptions before the number, and ask follow-up questions in the same chat.
  5. Run any SQL in your own warehouse and compare a few rows against the source before you report.

Example output

The question analysts ask most is whether a difference is real. This worked example runs a two-proportion z-test on an A/B test of a checkout page, then on the same rates with a tenth of the traffic.

Test 1: 2,400 visitors per variant
  A: 120 orders / 2,400 = 5.00%
  B: 156 orders / 2,400 = 6.50%
  Pooled rate p = 276 / 4,800 = 5.75%
  Standard error = sqrt(p x (1 - p) x (1/2,400 + 1/2,400)) = 0.00672
  z = (0.0650 - 0.0500) / 0.00672 = 2.23
  Two-sided p-value = 0.026  -> below 0.05

Test 2: same rates, 240 visitors per variant
  A: 12 / 240 = 5.00%    B: 16 / 240 = 6.67%
  Standard error = 0.0214
  z = 0.78, two-sided p-value = 0.44  -> no evidence of a difference

Both tests show B ahead, but only the larger one gives a result you can defend. A good analysis states the test, the sample and the p-value together, and says that a small p-value does not mean the lift will hold next month.

Limits and privacy

  • It writes SQL for Snowflake, BigQuery, and Postgres and other dialects you name, but it does not connect to your warehouse and run the query for you.
  • Python analysis on a file you upload runs in an isolated sandbox, so those results are computed. SQL meant for your own database is a draft until you run it.
  • A test describes your sample. It cannot tell you why a difference happened, and a significant result can still be too small to matter.

Data skills run like any other Zeplik chat, so a file you upload goes into that conversation and to the model you picked. Remove names, emails and other personal fields, or replace them with IDs, before you upload a customer export.

How can a SQL join change the answer without an error?

An inner join drops every row that has no match, so a count can shrink with no warning. The data analyst explains each join and filter in the query it writes, so you can see what it counts.

customers: 1,000 rows      orders: 640 rows from 410 customers

SELECT COUNT(*) FROM customers c JOIN orders o ON o.customer_id = c.id
  -> 640 rows: the 590 customers with no order are gone

SELECT COUNT(DISTINCT c.id) FROM customers c
  LEFT JOIN orders o ON o.customer_id = c.id
  -> 1,000 customers, 590 of them with no order

Average orders per customer:
  inner join view: 640 / 410 = 1.56
  all customers:   640 / 1,000 = 0.64

Both numbers are right for a different question. Say which one you mean before you ask for the SQL.

What should you give the data analyst for a good first profile?

Give it the file and one sentence on what each row is, such as "one row per order". The profile then reports column types, missing values, outliers and odd distributions against that grain, instead of guessing what a row means.

When is the coding assistant a better fit?

Use the AI coding assistant when the job is the application code around the data, such as an API endpoint or a migration script. The data analyst is for the question, the query and the statistics.

More on Zeplik

Ask your data a question

Upload a dataset or paste your schema and get the SQL, the analysis, or the dashboard you need. Free to start, and honest about which figures to check against the source before you report them.