readme.md

SENTINEL-CHECK API

Ensemble model fake review detection for small businesses.

Presidential AI Challenge Submission

THIS IS SOLELY FOR SUBMISSION TO THE PRESIDENTIAL AI CHALLENGE!

This is the deployed and final iteration of our submission. Any commits after the submission date are generally for maintenance purposes.

All documentation on the API can be found at the bottom of the "App" section of this space - look for the hyperlink button with the text "use via API".

---

Prediction Meanings:

• "genuine" - review is probably authentic
• "fake" - review is probably fake/fraudulent
• "uncertain" - models disagree (agreement < 67%); less than 2/3 models agree.

---

Model Arch

The API uses an ensemble of three models:

1. DistilBERT (distilbert-base-uncased) - max 128 tokens
2. RoBERTa (roberta-base) - max 192 tokens
3. BERT (bert-base-uncased) - max 256 tokens

Ensemble Strategy: Equal weighting (0.333 each)
Decision Threshold: 0.45 for fake classification
Uncertainty Threshold: 0.67 model agreement req.

Models loaded from HuggingFace Hub here is the link: https://huggingface.co/codingcoolfun9ed/sentinelcheck-models

---

Performance Metrics

• Test Accuracy: 78.26%
• Validation Accuracy: 84.64%
• F1 Score: 0.8531
• Model Agreement: 88.4%

Threshold optimized on validation set. Equal weighting outperformed an accuracy-based weighting approach.

---

Tech Stack

• Framework: Flask + Flask-CORS
• ML: PyTorch, Transformers (HuggingFace)
• Models: BERT, RoBERTa, DistilBERT
• Deployment: Python 3

---

Text Preprocessing

Reviews are automatically cleaned:

• URL removal
• HTML tag removal
• Punctuation normalization
• Whitespace normalization

---

Author

codingcoolfun9ed (Aaban R.) | Submission Date: January 20, 2026

For API documentation and the live demo, visit the app interface and/or click "use via API" at the bottom.