KaggleAgent.AI

The Grandmaster's Intelligence Hub
Built & Engineered by devparikh

"A small thing done well beats a big thing done badly."

KaggleAgent.AI is an autonomous, multi-agent intelligence platform designed to solve an obvious but painful problem: Kaggle Information Overload. For every competition, a data scientist must wade through hundreds of forum posts, dozens of code notebooks, and complex data documentation.

This agent mimics the exact sequential thought process of a Kaggle Grandmaster to distill that noise into actionable competitive intelligence in under 3 minutes.

---

🎯 Why This? (Taste & Judgment)

In building KaggleAgent.AI, the focus was on Distillation over Decoration.

• What’s Included: A strictly orchestrated 6-agent pipeline using LangGraph. Each stage (Overview → Data → Approaches → Winners → Forum → Code Scout) is dependent on the previous, ensuring the AI "reasons" through the competition rather than just summarizing it.
• What’s Left Out: We avoided generic chat interfaces as the primary entry point. Instead, we built a deterministic execution graph that guarantees a complete analysis every time.
• Smallest Interesting Version: The core "V1" focus was the streaming graph UI and the context-aware "Ask the Grandmaster" chat—providing a functional end-to-end mentor that works universally on any Kaggle URL.

🛠️ Originality & Architecture

Unlike basic LLM wrappers, KaggleAgent.AI uses a multi-agent state machine:

1. Overview Agent: Identifies technical "System Problems" (noise, distribution shifts).
2. Data Agent: Audits dataset architecture and evaluation metrics.
3. Winners Agent: Extracts the "Secret Sauce" from the Top 5 historical solutions.
4. Code Scout: Autonomously hunts for high-fidelity GitHub repos and research papers.
5. Grandmaster Chat: A context-aware expert that "lives" in the analysis, ready for deep-dive Python questions.

🚀 Shipping Ability (It Works)

The project is built to be used now.

• Real-time Streaming: Uses Server-Sent Events (SSE) to stream analysis results as they happen.
• Dynamic UI: A reactive dashboard with a visual execution pipeline.
• Export Ready: One-click professional Markdown report generation for offline sharing.

---

🏃 Quick Start (short & brief)

1. Requirements

Ensure you have Python 3.10+ and a .env in the /backend folder:

GROQ_API_KEY=your_key
TAVILY_API_KEY=your_key

2. Launch Backend

cd backend
pip install -r requirements.txt
uvicorn main:app --reload

3. Launch Frontend

Open frontend/index.html in any modern browser.

---

Built for the AI Build Challenge. Original work by devparikh.