Nikhil1417/Ann-Drift-detection

🔍 ModelShift Monitor

A lightweight, end-to-end data drift detection tool built with a pre-trained autoencoder, a FastAPI inference backend, and a Streamlit frontend.

Upload any CSV — ModelShift scores each row against the training distribution and tells you whether your live data has drifted.

🗂️ Project Structure

ModelShift-Monitor/
├── app.py                  # Streamlit UI
├── main.py                 # FastAPI backend
├── requirements.txt
├── runtime.txt
├── .streamlit/
│   └── config.toml         # UI theme
├── .devcontainer/
│   └── devcontainer.json   # VS Code dev container
└── model/
    ├── autoencoder_drift_model.h5
    ├── scaler.pkl
    ├── feature_columns.pkl
    └── drift_threshold.json

⚙️ How It Works

1. 1.Autoencoder — trained on the original (clean) dataset to learn its distribution.
2. 2.Reconstruction error — for each incoming row, the model attempts to reconstruct it. High error → the sample is out-of-distribution.
3. 3.Drift ratio — the fraction of samples whose error exceeds the calibrated threshold.
4. 4.Decision — if more than 30 % of samples exceed the threshold, drift is flagged.

🚀 Running Locally

1 — Install dependencies

pip install -r requirements.txt

2 — Start the FastAPI backend

uvicorn main:api --reload --port 8000

3 — Launch the Streamlit app

streamlit run app.py

Open http://localhost:8501 in your browser.

🌐 API Endpoints

Example request body

{
  "columns": ["feature_1", "feature_2"],
  "data": [[0.5, 1.2], [0.9, 0.3]]
}

Example response (/detect_drift)

{
  "drift_ratio": 0.12,
  "drift_detected": false
}

🧩 Tech Stack

📌 Notes

• —Input CSV must contain numeric columns that overlap with the training feature set.
• —Rows are capped at 1 000 before inference for latency reasons.
• —The hosted backend is on Hugging Face Spaces and may cold-start — allow ~30 s on the first request.