Arseny Afonin — Digital & AI for culture and education

Data that answers questions

At Soft Kultura, anyone on the team can use ChatGPT or Claude to test a hypothesis against our company data. They can ask a question in their own words, even when answering it takes several datasets and a statistical analysis they couldn’t run on their own.

(2026)
for my own company, Soft Kultura
one month from idea to rollout

The problem

We already used AI chatbots at work. We could upload a spreadsheet or connect a data source to an agent. But when a question involved several systems, the assistant had no reliable way to tell which records belonged together or how to interpret the figures.

Someone who knew the data had to explain those relationships and definitions in each new chat. The sources we shared could also contain customers’ personal details.

What I built

I built a data store that brings together figures on website traffic, leads, sales and company finances. It updates every night, matches related records across sources, checks the data and removes customers’ personal details before the team can query it.

Through MCP, ChatGPT and Claude can read both the data and a catalog of what our team knows about it. The catalog explains what each field means, how the datasets fit together and which figures are easy to misread. The assistant uses it to decide what to query.

The result

  • The CEO and every team lead can test hypotheses across several datasets and use statistical methods they couldn’t apply on their own.
  • The system prepares the data each night. The assistant then queries and analyses it for the person asking. That covers work we’d otherwise need a data engineer and an analyst to do.
  • Our dashboards use the same snapshot, so the team can check the figures in a chat against its regular reports.

How it works

Every night, a Python job collects data from our accounting database, CRM, ad accounts and web analytics. It aligns the records to the same point in time, matches them across sources and removes personal details. It saves the snapshot as Parquet files on AWS S3.

The MCP server loads the snapshot into DuckDB. The assistant reads the catalog and writes a SQL SELECT query. The server runs it with read-only access and limits both execution time and result size. Internal dashboards read the same snapshot through an API.

WorkOS manages access to the data served by the MCP server. The access log records who used the system and leaves out the questions they asked.