SemiGuard

ML-powered semiconductor defect prediction dashboard. Uses the UCI SECOM dataset (590 sensor features from real wafer fabrication) to predict pass/fail on wafers, with drift detection to catch when the model starts degrading.

Live demo: chipguardai.com

Stack: PyTorch, Flask, Angular, SQLite, AWS (Amplify + EC2 + ALB)

Architecture

+----------------+     +----------------+     +----------------+
|    Angular     |---->|   Flask API    |---->|    SQLite      |
|   Dashboard    |<----|   + PyTorch    |<----|   Database     |
+----------------+     +----------------+     +----------------+
                              |
                        +-----+-----+
                        |  Trained  |
                        |   Model   |
                        | (.pt file)|
                        +-----------+

• Frontend (Angular) - dashboard with stats cards, charts, drift monitoring, prediction form, and feedback UI
• Backend (Flask) - REST API serving predictions, logging to SQLite, drift detection via KS-test
• Model (PyTorch) - feedforward neural net trained on 562 features (after dropping high-missing columns)

Setup

1. Get the data

Download the UCI SECOM dataset from Kaggle and put the CSV in data/:

data/uci-secom.csv

2. Backend

cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

3. Train the model

cd backend
python train.py

This saves the model, scaler, and metadata to models/.

4. Run the API

cd backend
source venv/bin/activate
python app.py

API runs on http://localhost:5050.

5. Frontend

cd frontend
npm install
npx ng serve

Dashboard at http://localhost:4200.

API Endpoints

Method	Endpoint	Description
GET	/health	API status and model info
POST	/predict	Run prediction on sensor features
GET	/predictions	Recent prediction history
GET	/metrics	Prediction stats (pass/fail rates, counts)
POST	/feedback	Submit ground truth for a prediction
GET	/drift	Run drift detection on recent predictions
GET	/sample	Get a random data row for testing

Running Tests

cd backend
source venv/bin/activate
python -m pytest tests/ -v

Docker

docker-compose up --build

Backend on :5050, frontend on :4200.

What's going on with the model

• Feedforward net: 562 -> 128 -> 64 -> 1
• BCEWithLogitsLoss with pos_weight to handle the ~14:1 class imbalance
• Dropout (0.3) on hidden layers
• Trained for 50 epochs with Adam
• Drift detection uses Kolmogorov-Smirnov test per feature against the training baseline

Deployment

The app is deployed on AWS. Frontend and backend are separate.

Diagram

+----------------+     push to main     +-------------------+
|    GitHub      |--------------------->|  GitHub Actions   |
|   Repository   |                      |  (CI/CD)          |
+----------------+                      +-------------------+
                                           |            |
                               tests pass  |            | auto-trigger
                               + deploy    |            |
                                           v            v
                                    +----------+  +-----------+
                                    |   EC2    |  |  Amplify  |
                                    |  Docker  |  |  Angular  |
                                    |  Flask   |  |  Frontend |
                                    +----+-----+  +-----------+
                                         |              |
                                    +----+-----+        |
                                    |   ALB    |<-------+
                                    |  HTTPS   |  API calls
                                    +----------+

• Frontend is on AWS Amplify. Pushes to main auto-deploy.
• Backend runs in Docker on an EC2 t3.small instance. An Application Load Balancer handles HTTPS termination in front of it.
• Domain is chipguardai.com via Route 53. ACM provides the SSL certs.
• CI/CD is a GitHub Actions workflow that runs pytest, then SSHes into EC2 to pull and rebuild the container.