An autonomous multi-agent system executing the full machine learning research lifecycle.
- Finalist at both the LUMA Hackathon (top 3% of 397 teams) and the Open Source for AI Hackathon
- Agents: 19 specialized agents
- Research Phases: 8 end-to-end stages
- Pipeline Coverage: Full lifecycle (data → deployment)
- Execution Mode: Autonomous, multi-agent orchestration
- Output: Validated models + structured research artifacts + APIs
ML-Labs is a fully autonomous machine learning research system that replaces fragmented workflows with a unified, agent-driven architecture.
It was built for the Luma AI Agents Hackathon with a demo focused on training an interactive 3D graph neural network with competition datasets.
| Phase | Description |
|---|---|
| 1. Initiation & Planning | Defines objectives, constraints, and system scope |
| 2. Data Foundation | Ingests, structures, and validates datasets |
| 3. Model Development | Selects architectures and builds candidate models |
| 4. Validation & Testing | Runs automated tests and robustness checks |
| 5. Evaluation | Benchmarks performance and analyzes results |
| 6. Research Output | Produces structured reports and documentation |
| 7. Execution Layer | Builds systems, APIs, and deployment pipelines |
| 8. Continuous Loop | Iterates through feedback and refinement cycles |
| Agent | Role |
|---|---|
| Control Agent | Orchestrates workflow and resource allocation |
| Problem Framing Agent | Defines objectives, scope, and success metrics |
| Data Analysis Agent | Explores datasets and identifies patterns |
| Schema Agent | Designs data structures and transformations |
| Quality Agent | Ensures data consistency and reliability |
| Readiness Agent | Validates data for modeling |
| Model Selection Agent | Chooses algorithms and architectures |
| Model Agents (Ensemble) | Builds and iterates candidate models |
| Optimization Agent | Performs tuning and search |
| Testing Agent (Integrated) | Runs automated validation checks |
| Field Testing Agent | Evaluates models in real-world scenarios |
| Evaluation Agent | Measures performance vs objectives |
| Statistical Agent | Performs statistical analysis and inference |
| Full Stack Agent | Designs APIs and infrastructure |
| SWE Agent | Implements backend systems |
| Senior SWE Agent | Reviews code for scalability and reliability |
| System Testing Agent | Executes end-to-end validation |
| Computing Agent | Manages environments and execution |
| Research Agent | Generates structured outputs and reports |
- Sources: CSV, Kaggle datasets
- Type: structured and semi-structured tabular data
- Flow: ingestion → validation → feature construction → modeling
Preprocessing:
- schema construction
- normalization
- anomaly detection
- feature engineering
npm run devcurl -X POST http://localhost:3000/api/lab/runcurl -X POST http://localhost:3000/api/lab/predictInput: raw dataset Output: trained models + evaluation metrics + predictions
Dependencies
Node.js
Python 3.x
scikit-learn
NumPy
pandasGET /api/lab/demo
POST /api/lab/run
POST /api/lab/predict
/demo→ deterministic experiment snapshots/run→ full multi-agent execution/predict→ inference on trained models
ml-labs/
├── frontend/
├── backend/
├── agents/
├── pipelines/
├── data/
├── experiments/
└── README.mdnpm install
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txtnpm run build
npm start