📰 SwiftScan – Generative-AI News Detector

Lightweight browser extension + local API that flags paragraphs likely written by an LLM

---

1 . What SwiftScan does

Layer	Purpose
Chrome / Edge extension	Injects a content script into every page, splits visible text into paragraphs, sends each paragraph to a local REST endpoint (/predict), then overlays a yellow–red highlight if the AI probability ≥ 0.60.
Local FastAPI server	Loads a fine‑tuned DistilBERT classifier (directory model/) and returns a probability P(AI) for each request in < 30 ms on CPU.
Data/Training utilities	util/data.py → scrapes real‑world news, util/ai_data.py → generates GPT‑4o news, train.py (upcoming) fine‑tunes DistilBERT.

---

2 . Model at a glance

Item	Value
Base model	distilbert-base-uncased
Fine‑tuned on	~10 k balanced news chunks (AI vs real)
Sequence length	256 tokens
Metric (F1)	≈ 0.88
Files	model/config.json, pytorch_model.bin, tokenizer.json (≈ 45 MB INT8)

---

3 . Quick‑start

git clone https://github.com/yourname/swiftscan.git
cd swiftscan
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m spacy download en_core_web_sm
python server.py        # launches API on http://127.0.0.1:5000

Load the extension/ folder in chrome://extensions (Developer mode → Load unpacked).

---

4 . Configuration knobs

File	Variable	Meaning	Default
server.py	THRESHOLD	highlight if prob ≥	0.60
content.js	DETECTOR_ENDPOINT	REST target	http://127.0.0.1:5000/predict

---

5 . Data / retraining scripts

Task	Script
Scrape real news	util/data.py
Generate AI news	util/ai_data.py
Fine‑tune model	train.py (upcoming)

---

6 . Road‑map

• Bespoke news model
• Popup dashboard
• Image AI detection
• Optional cloud API

---

7 . Troubleshooting

Symptom	Fix
OPTIONS /predict 405	CORS middleware missing
ERR_CONNECTION_REFUSED	server not running
Slow pages	adjust limits in util/data.py

---

© 2025 SwiftScan Project – MIT License