MLOps · Pipeline Orchestration Python · Flask · Airflow · MLflow Live on HuggingFace Spaces

AutoMLOps — ML Experiment Tracking & Pipeline Studio

A full-stack MLOps platform that brings together visual DAG pipelines, automated machine learning, and experiment tracking in a single Docker container. 50+ scikit-learn, XGBoost, and LightGBM algorithms — orchestrated by real Apache Airflow DAGs and tracked with MLflow.

2025–2026 Mohammad Noorchenarboo 5 datasets · up to 20,640 samples 50+ algorithms · 2 task types
50+
ML algorithms across 7 categories
5
Built-in datasets ready to train
3
Pre-built Airflow pipelines
97.2%
Best accuracy (LightGBM on Wine Quality)
1
Docker container — no external services
Architecture Overview

End-to-End MLOps Architecture

AutoMLOps runs entirely within a single Docker container, starting an Apache Airflow Scheduler alongside a Gunicorn-served Flask API. Raw datasets from scikit-learn and real-world sources are preprocessed and fed to any of 50+ algorithms, with every run logged to an SQLite-backed MLflow tracking server. The Pipeline Studio renders live DAG execution state directly in the browser — no page reloads required.

🗄️
Datasets
sklearn + California Housing
🔧
Preprocessing
Scale · Encode · Split
🧠
Training
50+ algorithms via Airflow DAG
📈
MLflow
Params · Metrics · Artifacts
🚀
Registry
Staging → Production
💡

Single-Container MLOps Stack

The Airflow Scheduler starts first via start.sh, which polls until DAGs are parsed before launching Flask. All state (MLflow runs, Airflow metadata, model registry) persists in SQLite within the container — making the entire stack deployable to HuggingFace Spaces with zero external dependencies.

Module Breakdown

Six Core Modules

🎨 Studio
Pipeline Studio
Full-screen interactive DAG canvas rendered with HTML/CSS absolute positioning and SVG bezier arrows. Click any node to open a slide-in config panel; execution logs stream live in a built-in terminal.
Pipelines3 pre-built DAGs
EngineApache Airflow
🤖 AutoML
AutoML Engine
Automatically sweeps all algorithms in the registry for a chosen dataset and task type. Each trial is tracked as an MLflow run so you can compare results across the full algorithm landscape.
Search space50+ algorithms
Max runsconfigurable (default 20)
✈️ Airflow
Airflow Scheduler
Real Apache Airflow 2.10 with SequentialExecutor running inside the container. DAGs use PythonOperator and XCom to pass results between tasks, with DagRun/TaskInstance status surfaced back to the UI.
VersionAirflow 2.10.4
ExecutorSequentialExecutor
📈 MLflow
MLflow Tracking
Every training run — whether manual, AutoML, or pipeline — logs parameters, metrics, and sklearn model artifacts to a shared SQLite MLflow store. Experiments are organised by dataset name.
BackendSQLite (mlflow.db)
Trackedparams · metrics · artifacts
📦 Registry
Model Registry
Browse all registered models, their versions, and lifecycle stages (None → Staging → Production → Archived). Register any MLflow run directly from the UI and promote versions with a single click.
Stages4 lifecycle stages
APIMLflow Model Registry
🗂️ Datasets
Dataset Library
Five built-in datasets covering both classification (Iris, Wine, Breast Cancer) and regression (Diabetes Progression, California Housing). All expose a consistent loader interface used by training, AutoML, and pipelines.
Datasets5 built-in
Tasksclassification · regression
Algorithm & Tech Stack

50+ Algorithms Across 7 Categories

The algorithm registry covers the full spectrum from interpretable linear models to deep MLP networks. XGBoost and LightGBM are included as first-class citizens alongside scikit-learn. All algorithms share a uniform interface — the same JSON-serialisable config dict is used by manual training, AutoML search, and pipeline execution.

Linear Models
Logistic Regression (L1/L2), Ridge, SGD, Passive Aggressive, LDA · Ridge/Lasso/ElasticNet/Bayesian/Huber for regression
12 algorithms
Ensemble / Boosting
Gradient Boosting, AdaBoost, Bagging, XGBoost, LightGBM — for both classification and regression
10 algorithms
Tree-Based Models
Decision Tree, Random Forest, Extra Trees, QDA — classifier and regressor variants
7 algorithms
Neural Networks & SVMs
MLP (Small/Medium/Deep), SVC (RBF/Poly/Linear/LinearSVC), SVR (RBF/Linear), KNN k=3/5/9
17 algorithms
⚙️

Uniform Algorithm Interface

Every entry in ALGORITHMS carries class, params, description, and color keys. The algorithms_for_json() helper strips non-serialisable keys so the same registry drives the API, UI dropdowns, and AutoML search — zero duplication.

Interactive Explorer

Pipeline Scenario Explorer

Select a pipeline scenario to see representative metrics from the live platform. All values are from real training runs logged to MLflow.

Metrics from demo seed runs. Live app scores in real time — train your own models in the Pipeline Studio.

Performance Snapshot

Algorithm & Dataset Performance

Classification Accuracy
Dataset Distribution
Regression R² Scores

Top classification algorithm accuracy on demo datasets. LightGBM leads with 97.2% on Wine Quality.

Sample distribution across the 5 built-in datasets. California Housing is the largest at 20,640 samples.

R² scores for regression algorithms on California Housing. LightGBM Regressor achieves the highest R² of 0.834.

Design Decisions

Key Engineering Choices

✈️
Real Airflow
Pipelines execute as genuine Airflow DAGs with XCom, TaskInstance status tracking, and DagRun polling — not a simulated engine. The fallback is only activated if Airflow is not installed.
📦
Zero External Services
MLflow tracking, Airflow metadata, and the model registry all use SQLite. No Postgres, Redis, or message broker needed — the entire platform runs in a single HuggingFace Space container.
🎨
Canvas-Native DAG UI
The Pipeline Studio uses HTML divs with absolute positioning and SVG bezier arrows — no graph library dependency. Node status animations and the slide-in config panel are pure CSS transitions.