Why we built Creeklytics: an AI analyst with numbers you can check

01

The space between a spreadsheet and a decision

A useful question rarely arrives neatly packaged. Why did revenue fall last month? Which customers are waiting longer? Is a new process actually improving outcomes? The answer may be buried in an export, spread across sheets, or sitting in a company data system. Before anyone can decide what to do, someone has to turn those rows into a trustworthy explanation.

Spreadsheets make that work approachable. You can see the rows, change a formula, and follow a thought without raising a ticket. But bigger files, repetitive preparation, and increasingly complex questions can interrupt that flow. Reporting tools bring structure, yet a new question can still require a new report. Neither approach is wrong; the friction sits between them.

We built Creeklytics for that space. It brings a familiar spreadsheet together with a conversational analyst, so the next question can follow naturally from the last. Our aim is not to replace judgment with a confident paragraph. It is to shorten the distance between a question, a calculation, and a decision you can explain to someone else.

02

A fluent answer is not the same as a correct answer

An AI assistant can be excellent at interpreting a question and communicating a result. That does not make a generated percentage reliable. A rate needs a numerator, a denominator, and an agreed definition. If any of those change silently, the answer can look reasonable while meaning something entirely different.

Consider a fictional team with 120 enquiries and 30 purchases. The purchase rate is 25%. If an assistant excludes 20 unanswered enquiries without being asked, it reports 30%. Both divisions are arithmetically correct. Only one answers the original question. The problem is not simply arithmetic; it is the definition being used.

In Creeklytics, the Analyst requests calculations from the calculation engine rather than supplying its own mental arithmetic. Rate checks work from executed numerator and denominator results. A zero denominator has no meaningful rate, so it should not become a reassuring zero. Percentages without calculation evidence are marked Not verified instead of being presented as established facts.

These checks reduce important failure modes. They do not promise an infallible assistant. An ambiguous question, an incomplete source, or the wrong business definition can still lead to a misleading conclusion. That is why saved business metrics, clarification, and a visible result table matter. A number becomes useful when you understand what was counted—and what was not.

“Both divisions are correct. Only one answers the original question.”

03

Bring the work to the data

There are two practical paths through Creeklytics. For uploaded files, much of the calculation can happen on your own device. Large-file analysis uses a local query engine, so a summary does not require a remote computer to scan every row. This avoids cloud compute charges for that local calculation, and keeps the full file out of that calculation request.

Local does not mean unlimited. File size, column types, browser memory, and the device you are using all affect what runs comfortably. Opening a file and preparing it still take time. We would rather make those limits clear than call every operation instant. Saved large files can also be uploaded to private storage so they can be reopened; local calculation and cloud persistence are different choices.

For connected data systems such as Databricks and Snowflake, the better path is often to leave the full table where it already lives. Creeklytics asks the connected system to calculate a compact summary and returns bounded results. That avoids pulling an entire large table into a browser just to answer a question about a total, trend, or segment.

Your connected system still does the work, with its own permissions, availability, and billing. A paused warehouse can take longer to wake up than a running one. Connection setup is required, and no single performance claim fits every source. The principle is simple: calculate close to the data, then bring back the evidence needed for the decision.

04

Be clear about what is shared

Local calculation does not mean the whole experience is offline, or that nothing ever leaves your device. Conversational AI needs context. Depending on the workflow and sharing settings, that context can include column names, selected examples, and calculated summaries. A saved workbook or a saved large file also uses cloud storage.

Creeklytics includes controls for how rows are shared with AI and which columns are sensitive. Connections use protected credentials, and workbook access follows sharing permissions. Those controls work best when they are paired with deliberate choices: share only what is needed, review your sensitive columns, and use a connected account with appropriate permissions.

We describe these distinctions in our Data Usage and Privacy policies. Security is not a badge or a claim of certification we have not earned. It is a set of concrete boundaries, careful implementation, and honest explanations of what a feature does. If your team has stricter requirements, check those requirements before uploading or connecting confidential data.

05

The analyst should support judgment, not replace it

A useful analysis goes beyond a total. It helps you compare periods, examine segments, investigate what contributed to a change, and decide what to inspect next. Creeklytics brings these steps into a conversation, alongside charts and result tables that can be added back to a workbook.

But an association is not proof of a cause. A forecast is an estimate, not a commitment. A segment that looks unusual may reflect missing data rather than an opportunity. Good decision-making still needs subject knowledge and the willingness to question a plausible answer. The assistant should help you do that work, not disguise uncertainty.

“A faster answer you can responsibly use is better.”

06

Built by CreekData. Improved with you.

A creek is moving water: it finds a path, joins other streams, and keeps flowing. That is the idea behind our name. CreekData is the team; Creeklytics is our analytics product. We want the work to move smoothly from raw rows to understanding without hiding the steps in between.

Our background spans data engineering, analysis, data science, AI, and language technology. Creeklytics itself is new. We are starting with early testers, not borrowed endorsements or invented success stories. Their questions, bug reports, and feature requests help us discover where the experience needs to improve.

If you are exploring Creeklytics, start with data you are comfortable using for a test. Ask a question whose answer you already know, inspect the calculation, and then try a follow-up. Tell us what was useful and what was confusing. That feedback helps improve the product; it is not a request to send us your confidential records.

The ambition is straightforward: ask your data, understand the answer, and keep the evidence close enough to check. A faster answer is valuable. A faster answer you can responsibly use is better.