2025
AI/MLAutoML Agent
Built an LLM-orchestrated end-to-end ML pipeline that automates data preprocessing, model selection, hyperparameter tuning, and deployment with leaderboard visualization.
Overview
AutoML Agent is an autonomous machine learning system that ingests tabular datasets, cleans and preprocesses them, trains several candidate models, tunes hyperparameters within a fixed compute/time budget, and selects the best model against a target metric. The pipeline is LLM-orchestrated: an agent generates and refines Python code for each stage, executes it inside a controlled Docker runtime, and iterates until performance converges. A leaderboard UI displays metrics, model artifacts, and brief natural-language summaries.
Challenge
Building an autonomous ML system that can handle diverse datasets while maintaining safety and reproducibility required safely executing LLM-generated code without polluting the host environment, handling diverse data types with consistent preprocessing, optimizing models efficiently within a bounded budget, making results transparent and reproducible, and packaging outputs for reuse.
Solution
Developed a comprehensive AutoML solution using Docker as a sandboxed execution environment for the LLM orchestrator, supporting a curated model set (Linear/Logistic Regression, Random Forest, XGBoost) with random and Bayesian search via Optuna, implementing an auto-refinement loop for continuous improvement, and exposing a FastAPI backend with Next.js dashboard for leaderboard visualization.
Impact
Achieved a reliable, repeatable AutoML workflow with clean separation between the app layer and a Docker-based execution runner. Reduced environment drift and dependency issues while meeting compute budgets, improved best-model metrics through iterative refinement, and delivered a clear leaderboard and artifacts that enable immediate reuse for batch or single-instance inference.
Technologies
YEAR
2025
CATEGORY
AI/ML
