AgentHive

An AI agent ecosystem that automatically finds datasets, trains models, evaluates accuracy, and delivers ready-to-use ML packages—just from a single sentence.

Comment

Inspiration

We wanted to solve a fundamental problem in machine learning: building an ML model still requires too much manual effort, technical skill, and time. Tasks like dataset discovery, preprocessing, training, and evaluation often take hours—even for experts.

This inspired us to ask a bold question: What if AI could build ML models on its own?

AgentHive was born out of this idea:

An AI system that uses multiple agents to automatically find datasets, train models, evaluate them, and return ready-to-use results—all from a simple natural language request.

What it does

AgentHive is an autonomous multi-agent AutoML system.

A user can simply type:

“Build a plant disease detection model.”

And AgentHive:

Understands the request using a Planner Agent (powered by Gemini)

Searches Kaggle for a relevant dataset

Uploads it to Google Cloud Storage

Trains a PyTorch model

Evaluates accuracy

Generates a downloadable model bundle

Everything is handled by independent agents coordinated through Supabase.

We also built DermAsist, a Flutter app using one of the trained models to predict skin fungal and dermatological conditions. This demonstrates how AgentHive-generated models can be used in real-world applications.

How we built it

We designed AgentHive as a modular multi-agent system:

MCP Server + Gemini LLM to interpret natural language requests

Planner Agent to convert the request into ML tasks and dataset keywords

Dataset Agent to auto-fetch datasets from Kaggle and upload them to GCP

Training Agent using PyTorch to train models on cloud-hosted datasets

Evaluation Agent to compute accuracy and create exportable model bundles

Supabase to coordinate agent workflow via project status updates

Google Cloud Bucket for dataset and model storage

React + Flutter frontends for testing and mobile deployment

Every agent runs independently on FastAPI and communicates only through Supabase state changes, making the system highly scalable and fault-tolerant.

Challenges we ran into

Combining multiple agents into one smooth workflow

Ensuring all agents sync correctly using Supabase states

Handling large Kaggle datasets and uploading them to GCP

Integration difficulties between Node.js backend and Python agents

Managing cross-service authentication (Kaggle, Supabase, GCP)

Time-consuming model training on limited hardware

Real-time logging and debugging across distributed systems

Accomplishments that we're proud of

Built a fully functional AI-Agent AutoML pipeline

Integrated Kaggle → GCP → PyTorch → Supabase end-to-end

Created a live mobile app (DermAsist) using a VibeML-trained model

Achieved successful autonomous dataset detection, training, and evaluation

Designed a modular system that can easily scale to new agent types

Demonstrated a real-world medical application using our generated model

What we learned

How to design and orchestrate multi-agent systems

Cloud integration best practices (Supabase, GCP Buckets)

Building AutoML pipelines and training models automatically

Handling asynchronous workflows and distributed systems

Flutter integration with ML models

Proper debugging and logging for multi-service architectures

Importance of modular design in large AI systems

What's next for AgentHive

Add CSV, tabular, and time-series support

Create a Hyperparameter Tuning Agent

Add an Explainability Agent (Grad-CAM, confusion matrices)

Build a Model Deployment Agent (auto-generated prediction API)

Create a model marketplace where users can download agent-generated models

Extend to NLP, audio, and multimodal datasets

Improve real-time monitoring and dashboards for agent workflows

Built With

Share this project:

Updates