Invoice & Freight Machine Learning

This project contains two ML pipelines built on data/inventory.db:

• Freight Cost Prediction (regression): predicts freight amount from invoice dollars.
• Invoice Flagging (classification): predicts whether an invoice should be flagged for review.

It also includes a Streamlit web app to run both models from a UI.

Project Structure

.
├── app.py
├── data/
│   └── inventory.db
├── freight_cost_prediction/
│   ├── data_preprocessing.py
│   ├── model_evaluation.py
│   └── train.py
├── invoice_flagging/
│   ├── data_preprocessing.py
│   ├── modelig_evolution.py
│   └── train.py
├── inference/
│   ├── predict_freight.py
│   └── predict_invoice_flag.py
├── models/
│   ├── predict_freight_model.pkl
│   ├── predict_flag_invoice.pkl
│   └── scaler.pkl
└── notebooks/
    ├── predicting_freight_cost.ipynb
    └── invoice_flagging.ipynb

Tech Stack

• Python 3.x
• pandas
• numpy
• scikit-learn
• joblib
• sqlite3 (stdlib)
• streamlit

Setup

1. Create and activate a virtual environment.
2. Install dependencies.

python -m venv .venv
source .venv/bin/activate
pip install pandas numpy scikit-learn joblib streamlit

Or install from file:

pip install -r requirements.txt

Data Source

• SQLite DB: data/inventory.db
• Main tables used:
  • vendor_invoice (freight regression + part of flagging features)
  • purchases (aggregated for flagging features)

Model 1: Freight Cost Prediction

Path: freight_cost_prediction/

Feature/Target

• Feature: Dollars
• Target: Freight

Models Trained

• Linear Regression
• Decision Tree Regressor
• Random Forest Regressor

Selection Logic

• Evaluate MAE / RMSE / R2
• Best model chosen by lowest MAE in freight_cost_prediction/train.py

Train Command

Run from repo root:

python freight_cost_prediction/train.py

Output Artifact

• models/predict_freight_model.pkl

Model 2: Invoice Flagging

Path: invoice_flagging/

Labeling Rule

flag_invoice = 1 when either condition is true:

• abs(invoice_dollars - total_item_dollars) > 5
• avg_receiving_delay > 10

Else flag_invoice = 0.

Features Used

• invoice_quantity
• invoice_dollars
• Freight
• total_item_quantity
• days_po_to_invoice
• total_item_dollars

Preprocessing

• StandardScaler fit on train split and saved as:
  • models/scaler.pkl

Model Trained

• RandomForestClassifier with GridSearchCV

Train Command

Run from repo root:

python invoice_flagging/train.py

Output Artifacts

• models/predict_flag_invoice.pkl
• models/scaler.pkl

Inference Scripts

Freight Inference

python inference/predict_freight.py

This script loads models/predict_freight_model.pkl and predicts Predicted_Freight.

Invoice Flag Inference

python inference/predict_invoice_flag.py

Note: the current script is a basic template and may require alignment with the classifier feature set and scaler pipeline for production usage.

Streamlit Web App

Path: app.py

The app includes two tabs:

• Freight Model: input invoice dollars -> predicted freight
• Invoice Flag Model: input six engineered features -> predicted flag + confidence (if available)

Run App

streamlit run app.py

Then open the URL printed in terminal (usually http://localhost:8501).

Streamlit Cloud Deployment

• Ensure requirements.txt is present at repo root (already added).
• Ensure runtime.txt is present at repo root (already added, python-3.11).
• In Streamlit Cloud app settings:
  • Repository: this repo
  • Branch: master
  • Main file path: app.py

End-to-End Workflow

1. Train freight model:
   • python freight_cost_prediction/train.py
2. Train invoice flag model:
   • python invoice_flagging/train.py
3. Launch UI:
   • streamlit run app.py
4. Test predictions via web forms.

Common Issues

• FileNotFoundError for model/scaler:
  • Train model pipelines first so files are generated in models/.
• Import errors when running scripts:
  • Run commands from repository root.
• Streamlit command not found:
  • Install streamlit in active environment: pip install streamlit.

Future Improvements

• Add requirements.txt and pinned versions.
• Unify module naming (modelig_evolution.py typo can be renamed safely).
• Add unit tests for preprocessing and inference contracts.
• Add one consistent inference API for both models (CLI + batch input).
• Add model/version metadata and experiment tracking.