🏆 1st Place — GenAI APIs Track
LauzHack Mini-Hackathon 2025
Organized by @EPFL AI Team and @LauzHack

Overview • Why It's Different from ChatGPT • Features • Tech Stack • How It Works • Getting Started • Future Work • Team

🎥 Demo: Coming soon — the demo video will be added here once it's ready!

DataLive.ai is a next-generation AI assistant for data exploration and visualization. It combines automated preprocessing, intelligent code generation, and multimodal reasoning — all through a sleek and interactive Streamlit app.

Upload a .csv file, clean your data, ask natural language questions, generate plots, and even receive AI-driven insights from those plots. All computation is performed locally, keeping your data private while maximizing speed.

This project was built in under 24 hours during the LauzHack Mini Hackathon 2025, where it won 1st place in the Generative AI Track.

Unlike general-purpose tools like ChatGPT or Code Interpreter:

• Your data is never uploaded to the LLM → We extract only essential metadata (e.g., column names, types, number of missing values) and send that to the model. This:

  • Preserves data privacy
  • Avoids token overload on large datasets
  • Ensures compatibility with sensitive or proprietary files

• All code is executed locally → No need to copy/paste code; the app renders output instantly in the interface.

• Lightning-fast interactions thanks to Streamlit → Near-instant feedback loop between code generation, execution, and visualization.

• Data shape, summary statistics
• Missing values report
• Column data types and inferred structure

Choose from:

• Drop rows with missing values
• Fill with column-wise mean
• KNN Imputation

• Powered by OpenAI GPT-4o

• Ask natural questions like:

  "Show a pairplot of numerical features" "Plot the class distribution as a bar chart"

• Returns executable Python code, auto-rendered as:

  • Seaborn heatmaps
  • Histograms
  • Boxplots
  • And more...

• After generating a plot, ask:

  "What can you conclude from this?"

• The image is sent to Video-LLaMA2B via a custom Hugging Face inference handler

• Returns textual interpretation of trends, outliers, and insights

• Save all generated code
• Export plots for reuse or reporting

• Frontend: Streamlit

• Backend: Python (OpenAI API + Hugging Face Inference Endpoint)

• LLMs:

  • GPT-4o (OpenAI) – Prompt-to-code generation
  • Video-LLaMA2B (DAMO-NLP-SG) – Plot interpretation

• Libraries: pandas, NumPy, Scikit-learn, Seaborn, Matplotlib

• Execution: Local sandboxed Python runtime (safe, fast, private)

1. Upload your .csv dataset

2. Receive an automatic summary of key stats

3. Select a data cleaning strategy

4. Ask a question or request a visualization

5. Get back both:

   • The generated code
   • The rendered output

6. (Optional) Ask for AI-based interpretation of the plot

pip install -r requirements.txt

streamlit run app.py

You'll need:

• OPENAI_API_KEY
• HUGGINGFACE_TOKEN (for the Video-LLaMA endpoint)

Here are some exciting features planned for the next iteration:

• SQL Data Integration → Connect to relational databases (PostgreSQL, MySQL) and run natural language queries directly on live data.

• NoSQL Support (MongoDB, Firebase, etc.) → Enable compatibility with document-based datasets.

• Advanced Insight Generation → Incorporate models to automatically detect:

  • Trends
  • Seasonality
  • Anomalies
  • Correlations and causal patterns

• Explainable AI Layer → Provide rationale behind detected trends or statistical recommendations.

Built by Ylan Vifian , Youcef Amar and Arno Vifian .