Agentic AI for visual data exploration

For those who might be interested, I’m sharing an output from my experimentation with agentic AI frameworks like LangChain and LangGraph in the context of data insights democratization.

It’s a simple app for plain-language exploration of a dataset from CultureX, containing corporate culture values across companies from different industries, measured through anonymous Glassdoor reviews between Jan 1, 2023, and Apr 4, 2025.

It consists of several agent-like components (powered by the GPT-5-mini model), each responsible for a different aspect of the insight-generation process:

• one component plans the analysis steps based on the user’s request, data characteristics, and broader context,
• another evaluates the quality of the proposed plan and suggests adjustments if necessary,
• another turns this plan into runnable code,
• another fixes potential errors in the code based on received error messages,
• another creates code for accompanying analyses to support explanation of the dataviz, and
• another transforms the results of these analyses into a concise, clear narrative for a non-technical audience.

You can see it in action in the attached video below, but feel free to try it out with your own questions here. (Note: Since the app is hosted on Streamlit Community Cloud, it doesn’t stay awake continuously. If it hasn’t been used recently, you may need to wake it up and wait a few minutes.)

It usually provides good outputs even when you use really simple plain language - but based on my experiments, it helps if you know what kind of dataviz and insights you want and spell that out for the agents 😉

Happy exploring 🕵️‍♀️

P.S. As usual, genAI can make mistakes, so don’t trust the outputs blindly - always double-check 😉

P.P.S Click the GitHub icon in the top right corner of the app to get to the code repo, or go directly here.

---

Update: After Google’s release of Gemini 3 Flash Preview, which demonstrates a superior performance-to-price ratio, I replaced GPT-5-mini with this model. Its higher “intelligence” is readily apparent in the quality of the app’s outputs - specifically through better chart selection and more nuanced data interpretations. It’s a perfect example of the ongoing commoditization of “intelligence.”