Ajeet Kumar

Ajeet Kumar

Applied Mathematician with Specialization in Full Stack Machine Learning and Data Science

AutoResearch: Autonomous ML Research Framework

Published:

🔬 AutoResearch Overview

AutoResearch is an agents and device-agnostic autonomous machine learning research framework that enables AI agents to independently conduct end-to-end experiments. Unlike traditional AutoML tools, AutoResearch provides a complete research loop where agents can modify code, run experiments, evaluate results, and learn from a persistent semantic memory.

🌟 Key Features

  • 🤖 Advanced Agent Integration: Built-in protocols for advanced agentic systems like Claude Code, OpenCode, and OpenClaw.
  • 🧠 Semantic Research Memory: Long-term RAG-based memory using ChromaDB allows agents to learn from every past experiment.
  • 📈 Industrial Observability: Native integration with Weights & Biases (W&B) for real-time metric tracking and artifact management.
  • ⚡ High-Performance Backends: Native support for Apple MLX, JAX, and PyTorch (CUDA/MPS/CPU).
  • 🚀 Distributed Scaling: Parallelize research across clusters using Ray.
  • 🛡️ Robust Orchestration: SQL-backed metadata management and automatic Git-based experiment versioning.

🏗️ Architecture

AutoResearch acts as the Research Platform (The Orchestrator) while the AI system acts as the Brain (The Agent).

🚀 Quick Start

Clone and navigate to the directory then:

1. Install

pip install .

2. Initialize a Project

autoresearch init my_research --name "Optimizing-Transformer"
cd my_research

3. Run with an External Agent (e.g., Claude Code) Configure your autoresearch.yaml:

agent:
  type: "external"
  command: "claude-code"

Then start the autonomous loop:

autoresearch run

🛰️ Autopilot Mode (Pro)

For high-throughput research, AutoResearch supports a fully autonomous “Autopilot” mode. This allows the framework to automatically drive interactive CLI agents like OpenCode or Claude Code by feeding them prompts and auto-exiting sessions once the code is optimized.

📊 Comparison: AutoResearch vs. Traditional AutoML

FeatureTraditional AutoMLAutoResearch
ScopeHyperparameter tuning onlyFull code & architecture modification
Agent ControlFixed search spaceAI decides what to change
LearningGrid/Bayesian searchSemantic memory (RAG) of past results
Device SupportVaries by toolNative MLX, JAX, CUDA, MPS
IntegrationLimited to configsDirect integration with Claude Code/GPT

👥 Team Roster for ajeetkbhardwaj/automlresearch

  • 👑 Team Leader:
    • ajeetkbhardwaj (Active Commits)
  • 👨‍💻 Team Members:
    • Open-Source Contributors (Active Commits)

📅 Weekly Plan & Updates

👑 Team Leader Update (Ajeet Kumar)

  • Solved: Configured autopilot: true in the autoresearch.yaml file so the framework can automatically drive the agents unattended.
  • Completed: Implemented the long-term semantic memory (RAG) using ChromaDB, allowing the agents to reference previous runs.
  • Next Steps: Optimize the SQL-backed metadata management for faster Git-based experiment versioning.

👨‍💻 Team Member Updates

  • Solved: Fixed the Weights & Biases (W&B) integration for real-time metric tracking.
  • Working on: Parallelizing the research environments across clusters using Ray.