When people hear "AI data analyst" they often picture a chatbot reading a spreadsheet and replying with numbers. We built PowerDataChat around a different loop:
question → AI model → Python code → sandboxed execution on the data → structured result → chat UI
Here is the loop at a high level.
1. Read the structure first. When a file is uploaded or a database table is connected, PowerDataChat reads its structure (tables, columns, types) and drafts plain-language column descriptions that a person can edit. Good descriptions make every later step better.
2. Plan and write code. The user asks a question in English, Georgian or Russian. The model gets the question, the structure, the descriptions and the chat history, and writes pandas code.
3. Run it somewhere boring and safe. The code runs in an isolated sandbox: no internet, restricted imports, time and memory limits. Boring is the goal.
4. Render something useful. Results come back as interactive Plotly charts, tables, statistics and a short written explanation. "Show code" and "Show data" sit behind every answer.
Why not let the model answer directly? Because the model is good at deciding what to compute and a Python runtime is good at computing it. Splitting the two makes answers checkable and repeatable, and when a needed column does not exist the honest answer is "that data is missing", not a made-up number.
Try the web version on your own CSV at powerdatachat.com (the Basic plan is free). Curious about a part of the loop? Ask in the comments.
Top comments (0)