LeanAI 2026

Predicting Body Fat Percentage Using Machine Learning

An end-to-end MLOps project that predicts body fat percentage from anthropometric measurements using SVR, achieving < 1% error (MAE: 0.10, R²: 0.9996).

---

Results

Model	MAE	R²	MAPE
SVR (best)	0.10	0.9996	0.86%
MLPRegressor	0.24	0.9979	1.82%
LinearRegression	0.50	0.9925	3.79%

Full benchmark (11 models):

SVR                0.103    0.9996    0.87%
MLPRegressor       0.248    0.9979    1.82%
StackingRegressor  0.319    0.9947    3.45%
LinearRegression   0.506    0.9926    3.79%
Ridge              0.525    0.9920    4.00%
GradientBoosting   1.315    0.9469    9.06%
XGBoost            1.502    0.9134    9.75%
RandomForest       1.703    0.9104   12.77%
Lasso              2.071    0.8885   16.67%
AdaBoost           2.291    0.8571   17.80%
ElasticNet         2.861    0.7875   23.82%

---

Tech Stack

Layer	Tool/Framework
Language	Python 3.12
Data Processing	Pandas, Polars, Scikit-learn
ML Models	SVR, XGBoost, MLP, Ridge, etc.
Feature Engineering	PolynomialFeatures, RFE, PCA
Workflow Orchestration	Metaflow
Experiment Tracking	MLflow
Model Monitoring	Evidently AI
HPO	Optuna
Feature Store	Featureform (YAML)
API	FastAPI
Frontend	Streamlit
Containerization	Docker + Docker Compose
Infrastructure	OpenTofu, K3s/Minikube
CI/CD	GitHub Actions
Linting	Ruff

---

Quick Start

Local Development

cd Project

# Install dependencies
pip install -r requirements.txt
pip install -e ".[dev,mlops,viz]"

# Train models
make train

# Run API
make api
# -> http://localhost:8000 (web form)
# -> http://localhost:8000/docs (Swagger UI)

# Run tests
make test

Docker

cd Project
docker compose up --build    # Start API on port 8000
docker compose --profile dev up -d   # Include Jupyter
docker compose --profile train run --rm train  # Train

API Usage

curl -X POST "http://localhost:8000/predict/" \
     -H "accept: application/json" \
     -H "Content-Type: application/x-www-form-urlencoded" \
     -d "abdomen=85&hip=100&weight=75&thigh=60&knee=38&biceps=32&neck=37"

---

Project Structure

Project/
├── api/                    # FastAPI prediction service
│   ├── main.py             # Endpoints: GET /, POST /predict/, GET /health
│   └── templates/form.html # Web form UI
├── modeling/               # ML training & inference
│   ├── train.py            # SVR pipeline: PolyFeatures -> RFE -> PCA -> SVR
│   └── predict.py          # Batch & single prediction
├── dataset.py              # Data loading, feature engineering, outlier removal
├── plots.py                # EDA visualizations (distributions, heatmaps, boxplots)
├── config.py               # Path configuration
├── tests/                  # pytest test suite
│   ├── test_api.py
│   ├── test_dataset.py
│   └── test_modeling.py
├── mlops/                  # MLOps pipeline
│   ├── src/flows/          # Metaflow workflows
│   ├── src/monitoring/     # Evidently drift detection
│   ├── src/retraining/     # Automated retraining flow
│   ├── features/           # Featureform YAML definitions
│   ├── dashboards/         # Streamlit dashboards
│   └── infra/              # Terraform/OpenTofu IaC
├── notebooks/              # Streamlit app & analysis
├── docker/                 # Dockerfiles (API, Jupyter, Train)
├── docker-compose.yml      # Service orchestration
├── Makefile                # Dev commands
├── pyproject.toml          # Project config, deps, ruff, pytest
├── pixi.toml               # Conda environment (cross-platform)
└── requirements.txt        # Core pip dependencies

---

Dataset

• Source: Kaggle - Body Fat Prediction
• Size: 436 samples, 16 features
• Target: Body fat percentage
• Features: Age, Weight, Height, Neck, Chest, Abdomen, Hip, Thigh, Knee, Ankle, Biceps, Forearm, Wrist, Density

Feature Engineering

• bmi = Weight / (Height/100)^2
• waist_to_hip = Abdomen / Hip
• waist_to_height = Abdomen / Height
• arm_ratio = Forearm / Biceps

---

ML Pipeline

graph LR
  A[CSV Data] --> B[Feature Engineering]
  B --> C[Outlier Removal<br/>z-score]
  C --> D[PolynomialFeatures<br/>degree=2]
  D --> E[RFE<br/>8 features]
  E --> F[PCA<br/>5 components]
  F --> G[SVR<br/>C=10, rbf]
  G --> H[MLflow Tracking]
  G --> I[Evidently Reports]
  G --> J[FastAPI Serving]

Loading

Models Trained

• Combined (all data)
• Male-only subset
• Female-only subset

Monitoring

• Data drift detection (Evidently)
• Target drift detection
• Regression performance reports
• Automated retraining via Metaflow when drift is detected

---

Development

# Lint & format
make lint
make format

# Run full test suite
make test

# Generate EDA plots
make plots

# Process dataset
make data

---

Architecture

graph TD
  subgraph Client
    A1[Web Form] --> A2[FastAPI]
    A3[REST Client] --> A2
  end
  A2 --> B1[ML Model<br/>joblib]
  B1 --> B2[MLflow Tracking]
  B1 --> B3[Evidently Monitoring]
  
  subgraph MLOps
    C1[Metaflow Orchestration]
    C2[Drift Detection] --> C3[Auto-Retrain]
    C4[Optuna HPO]
  end
  
  subgraph Infrastructure
    D1[Docker Compose]
    D2[GitHub Actions CI/CD]
    D3[OpenTofu IaC]
  end

Loading

---

Team

Team Member	Email
Alejandro Castellanos	k2jac9@users.noreply.github.com
Anna Wong	annawong.qea@gmail.com
Faisal Khan	fa.khan@alumni.utoronto.ca
Hassan Saade	saadehassan@hotmail.com
Igor Bak	baxwork88@gmail.com

UofT DSI Team-17

---

License

BSD License - see LICENSE