I built Dekalytics for Deka Insight to automate product-test reports.

In a typical test, hundreds of people try two products, such as two coffee recipes, and answer a survey. An analyst then cleans the data, runs statistical tests, and builds a 30 to 50 sheet Excel report for the client. At Deka, that used to take 2 to 3 days. Dekalytics does it in a few minutes.

From raw files to a report

Product-test reports follow a familiar pattern, but the details change with every study. One uses a 5-point scale, another uses 7 points, and a client may ask for a new demographic split after seeing the first report.

Traditional software needs every rule and edge case handled in advance. For Dekalytics, each new study could mean more fields and custom logic.

An AI agent is more versatile: it inspects the files, adapts the analysis and Excel layout, and revises both when the request changes. Researchers upload the questionnaire, raw data, and Form Processing workbook they already use, then explain anything unusual in English or Bahasa Indonesia.

The agent reads the files, detects the study design, rating scales, and demographic segments, then shows the researcher its analysis plan. If something is missing or ambiguous, it asks before running the numbers.

Once the plan is approved, it writes and runs Python to produce report.xlsx, including significance tests, crosstabs, driver analysis, and JAR tables. Researchers can inspect the workbook in their browser, request changes in chat, or share the finished report with a client without exposing the chat and source files.

What I had to get right

Fast enough for a conversation

I first used hosted E2B sandboxes, but waiting 10 to 15 seconds for a container to start felt slow in a chat. I moved the app and its Python runtime onto one VPS, so a project is ready as soon as the researcher opens it.

Familiar reports, correct statistics

I tested the output against reports made by Deka's analysts. In one early comparison, the significance results did not match. The agent used paired t-tests because the same people had tried both products; some older files treated them as independent groups. The agent was right for that study design.

The workbook still had to follow Deka's existing format, from Top Box and Just About Right tables to the table of contents. Correct numbers would not save much time if researchers still had to rebuild 40 sheets by hand.