Data you can’t hand to a cloud AI?

The local AI that gives you a first look at your data.

Your files hold answers, and most of them are hard to reach. The data tools want a data team, and the cloud AIs want your data in their cloud. With Totsum, the AI runs on your machine, next to your files, and you ask it questions in plain language.

We also believe that a person should check the work of any AI, so we built Totsum around a human check. The answers you keep go into a report, which your reviewer verifies and signs.

Totsum requires a GPU with roughly 12 GB of VRAM, or a Mac with an M4 Pro chip or better (an M3 Max or M3 Ultra also qualifies) and at least 18 GB of unified memory. Please check compatibility before downloading. We would rather you learn this now than after a 7 GB download.


How it works

Ingest. Explore. Validate.

Totsum works as one loop. You add your files, you ask them questions, and the loop ends with a report that your reviewer checks.

FIG. 1 The loop runs through the three panes of the app. You ingest a file, you explore it with questions, and you validate the results you shortlisted through a report.

“Ingest” is the data world’s word for bringing a file into a tool. You drop a file in, and Totsum reads it into tables that you can question. The file itself stays where it is, unchanged.

It explores your data, question by question.

The first hours with a new file cost the most. You don’t know yet where to look, and with the usual tools every idea you want to test means building an export, a pivot table, a formula or a chart. In Totsum you ask a question, read the answer and ask the next one, so testing an idea takes minutes instead of an afternoon.

The kind of questions

  • “Which ten products drive most of the returns?”
  • “How did weekend sales compare to weekdays, month by month?”
  • “Which country’s customers spent the most?”Even when the customers and the invoices live in two different files, Totsum looks for the relevant tables and joins them.

The kind of answers

With each answer you get a plain-language explanation, the chart, a preview of the rows behind it, and the query that produced them. When Totsum notices that it had to make an assumption in order to answer, such as which file, which column or which spelling to use, it says so with the answer, and you can correct it. You read the result, then you shortlist it or rule it out, and you ask the next question.

The AI Assistant panel in Totsum: the question "Which ten products drive most of the returns?" answered in plain language with a link to the chart, and a "Try asking" list of further questions.

Except when it is too complex, the generated SQL is also translated into a query flow, which gives a step-by-step account in plain language of what was read, filtered, grouped and calculated.

The AI runs on your machine.

The files that need AI the most are the ones you can’t hand to a cloud AI, such as client data under NDA, personal data, or anything a contract says must stay in-house. With Totsum, the language model and the SQL engine run on your own computer, and Totsum opens your file where it already is. That is also why Totsum needs a powerful machine.

Demo

Watch it work with the connection off.

In the demo, the network is switched off on screen, and then Totsum gives a confidential file its first look. You see the answer, the chart and the rows behind it appear.

The network is switched off first, and then the preview appears. Totsum can work on a computer with no network connection.

The rhythm

It works while you work.

An AI model that runs on your own machine needs time to work. Expect a few minutes for an answer rather than a few seconds, and Totsum is designed around that wait. With the usual tools, you stay at the keyboard for as long as the search takes. With Totsum, you ask the question, continue with your other work, and find the answer ready later, with the chart, the preview and the rows behind it. The demo above shows the wait.

Privacy

Your data is processed on your machine.

Totsum processes your datasets, questions, generated queries and results on your computer. Where the confidentiality of your work requires it, Totsum can work on a computer with no network connection.

FIG. 2 The AI model and the SQL engine run inside your computer, where your file is. There is no account, no sync, no telemetry and no update check.
  • Local AI model. The language model runs in memory on your machine rather than on a cloud API.
  • Local SQL engine. Every query runs in an embedded database engine on your computer.
  • No account, no sync, no telemetry, no update check. There is nothing to sign up for, and your license is verified on your device.
  • Can work on a computer with no network connection. The installer downloads the AI model once (~7 GB). On a machine with no internet access, the model files can be provisioned by hand instead, and the machine can stay disconnected.

The handoff

A report built for the human check.

AI is helpful, and that is what makes it risky. After a few good answers in a row from any AI, it is natural to relax and stop checking. Yet some answers will be wrong, and that is a certainty. When a number is wrong, the decision belongs to a person. This is why the report Totsum produces is a follow-up report and not a final one. It is addressed to the expert or the team who does the last check, and it carries what they need to check the work. It holds the question exactly as you asked it, the chart, a subset of the data behind it, the query flow and the generated SQL. It also has a verdict field for each result, and each verdict returns with your reviewer's name on it.

Page one of a follow-up report with a company letterhead: the report basis, the reviewer’s name, the original question, a note for the reviewer, a Valid verdict ticked, and the revenue-by-country chart.

It says what it is

The report opens with a statement of what it is, in these words: “These results come from exploring source data with AI assistance. They are a first look rather than a validated analysis: each one is recorded here for a qualified reviewer to judge, and the verdicts against them are that reviewer's.”

Your reviewer completes it

The exported report has fields to complete. Against each result, your reviewer records Valid, Invalid or Inconclusive, and the verdict carries the reviewer's name.

The reviewer is the person you choose to check the numbers before anyone relies on them: a colleague, your accountant, an analyst, or you, if that is your competence. The reviewer is not an AI, and not someone Totsum supplies.

The verdicts return to the app

When you import the completed review, each verdict appears in the app beside the result it judges. You keep a record of what was checked, by whom, and what they concluded.

Totsum never marks a result Valid. We built it so that the judgment belongs to a person and is recorded with that person's name, because a number that nobody has signed is a liability. An Invalid verdict is useful too, because understanding why an answer was wrong usually teaches you something about the data, such as what a column actually holds, or which assumptions the dataset makes without saying so. What Totsum can't do is written down too.

One-time purchase

You pay once, and the license is yours.

Most data tools today are metered, and they charge per seat, per month or per question. Totsum is sold under a one-time license. Because the AI runs on your hardware, there is no cloud bill per question for us to charge you for. You have no subscription to cancel, and the app you bought keeps working on your machine.

  • The license is perpetual. It costs €249 for Personal or €499 for Commercial, and it is sold in the European Union (your currency and tax are set at checkout).
  • Security and conformity updates for your major version are free during its support period.
  • There is no monthly charge and no seat renewal.

Compare plans

Before you join the beta

Totsum runs its AI model on your computer, next to your files, and the model needs a lot of memory there. That is why the hardware requirement is high, and most current computers do not meet it. See the hardware compatibility guide to check yours.

MacApple Silicon, M4 Pro or better (or M3 Max, M3 Ultra)18 GB unified
WindowsDedicated GPU — NVIDIA, AMD or Intel12 GB VRAM
AI modelDownloaded during installation, runs on your machine~7 GB on disk
NetworkCan work on a computer with no network connection
AI outputA first look; verdicts are recorded by your reviewer, never by the AI

We are running a private beta. A beta is the step before an app goes on sale. A small group of early users runs it first, free of charge, and the feedback they give shapes the finished product. We call it private because you join by applying rather than by downloading.

If your machine qualifies, apply with your email. We invite every applicant who lives in the European Union, and the invitation follows the confirmation email.

Apply for the private beta