pgalko/BambooAI

What it solves

BambooAI is an LLM-driven data analyst designed for professionals who require transparent, reproducible, and verifiable data analysis. Unlike standard chat interfaces, it operates as a structured analyst that writes and executes Python code in a persistent kernel, ensuring that every final result is re-executed and verified in a fresh environment before being reported to the user.

How it works

The system operates through a sequence of "turns" where an analyst model reads the current state (question, dataset, and previous cell outputs) and takes specific actions such as running a Python cell, searching the web, or requesting more information.

Key technical mechanisms include:

• Persistent Kernel with Checkpoints: Analysis happens in a Python kernel where each cell is a checkpoint; failed cells are rolled back to maintain a clean namespace.
• Verification Loop: After a report is generated, the system assembles the cited cells into a script and runs them in a fresh kernel to verify that the numbers in the report are reproducible.
• Multi-Seat Architecture: Different roles (Analyst, Reviewer, Rewriter, Knowledge Distiller) are assigned to different models based on the required reasoning effort or cost.
• Isolated Execution: By default, model-generated code runs inside a Docker container to ensure security and isolation from the host machine.
• Memory Packs: The system distills lessons learned about a specific dataset into a memory pack for use in future analyses of the same data.

Who it’s for

It is built for data analysts and researchers who need a step-by-step notebook-style workflow where intermediate results are visible and final numbers are strictly checked.

Highlights

• Reproducibility: Automatically re-runs cited code in a fresh kernel to validate report figures.
• Flexible Execution Modes: Offers Quick, Deep, and Adaptive modes (the latter featuring periodic self-review by a stronger model).
• Broad Model Support: Compatible with ten providers including OpenAI, Anthropic, Google, Groq, Mistral, xAI, DeepSeek, and local options via Ollama and vLLM.
• Integrated Data Sources: Native support for CSV, Parquet, JSON, XLSX, and specialized integrations for sports science data (Intervals.icu, Endura, SweatStack).
• Web Grounding: Uses Gemini for web searches to compare internal data against published figures.