Automated Scientist
AI-Powered Experimentation
Accelerate scientific discovery with AI agents that autonomously design, implement, and evaluate experiments – all while you watch in real-time.
Current Demo: Self-Modeling Transformer for MNIST
Watch as our AI agents design, implement, and evaluate a transformer model that can classify MNIST digits while simultaneously modeling its own attention mechanisms.
AI Agent Workflow
Our AI workflow orchestrates four specialized agents to automate the scientific experiment process for transformer model development
Design Agent
Analyzes experiment requirements and creates detailed specifications based on goal understanding, producing an implementation blueprint for our self-modeling transformer.
Implementation Agent
Converts design specifications into executable PyTorch code, implementing transformer architecture with self-modeling capability and establishing a complete experiment pipeline.
Validation Agent
Verifies code correctness, checking data loading, model architecture, optimizer configuration, and ensuring technical requirements like GPU compatibility are met.
Evaluation Agent
Runs the experiment, monitors training metrics, and produces visualizations of model performance including accuracy comparisons and loss curves.
View Experiment Demo
Watch a replay of AI agents designing, implementing, validating, and evaluating a transformer model with self-modeling capabilities.
Implement a transformer-based model with self-modeling for MNIST classification
Press "Play Demo Replay" to start