AI text detection and paraphrasing tool for academic papers. Detects AI-generated content using a calibrated ensemble of linguistic and statistical signals, then rewrites flagged sections to read as authentically human-written.

• Detection — Ensemble scorer combining LLM judge + 6 calibrated heuristic/linguistic features (burstiness, vocabulary markers, n-gram uniformity, paragraph structure, repetition, punctuation diversity)
• Paraphrasing — Domain-aware rewriting with iterative refinement, semantic similarity verification, and post-processing pipeline
• Document support — PDF, DOCX, LaTeX (.tex), plain text
• Reports — Turnitin-style annotated PDF reports with color-coded paragraph scoring
• API — FastAPI REST API with background job processing for long-running tasks

• Python 3.11+
• Redis (optional, falls back to in-memory)

git clone <repo-url> && cd rewrite
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Edit .env with your Pollinations API key

# Detect AI content in a file
python run.py detect -f paper.pdf

# Detect with per-segment breakdown
python run.py detect -f paper.tex --segments

# Generate a detection report
python run.py detect -f paper.pdf --report report.pdf

# Full pipeline: detect -> paraphrase -> report
python run.py process -f paper.tex -d cs

# Paraphrase text directly
python run.py rewrite "Your text here" -i aggressive -d cs

# Interactive mode
python run.py interactive

# Start the API server
uvicorn app.api.app:app --host 0.0.0.0 --port 8000

# Detect AI content
curl -X POST http://localhost:8000/api/detect \
  -H "Content-Type: application/json" \
  -d '{"text": "Your text here...", "segments": true}'

# Upload a file for detection
curl -X POST http://localhost:8000/api/detect/file \
  -F "[email protected]"

# Start a paraphrase job (returns job_id)
curl -X POST http://localhost:8000/api/paraphrase \
  -H "Content-Type: application/json" \
  -d '{"text": "Your text here...", "domain": "cs", "intensity": "aggressive"}'

# Poll job status
curl http://localhost:8000/api/job/{job_id}

# Download report PDF
curl http://localhost:8000/api/report/{job_id}/pdf -o report.pdf

# Health check
curl http://localhost:8000/api/health

docker compose up --build

API available at http://localhost:8000. Interactive docs at http://localhost:8000/docs.

• PDF (.pdf)
• Microsoft Word (.docx)
• LaTeX (.tex)
• Plain text (.txt, .md)

Max file size: 10 MB.

Ensemble of calibrated features (weights based on Cohen's d effect sizes from arXiv paper analysis):

Score thresholds: <20% human, 20-60% mixed, >60% likely AI.

Paraphrasing preserves domain-specific terminology:

• cs — Computer science (algorithms, complexity, system names)
• medicine — Drug names, dosages, p-values, clinical terms
• law — Case citations, statutes, legal terms
• humanities — Direct quotes, author names, page references
• general — No domain-specific preservation

• General endpoints: 60 requests/minute per IP
• Paraphrase endpoints: 5 requests/minute per IP (expensive LLM calls)

app/
  core/         Config, LLM wrapper, embeddings
  detection/    Heuristics, linguistic analysis, LLM judge, ensemble
  paraphrase/   Prompts, rewriter, post-processing, targeted rewrite
  document/     PDF/DOCX/LaTeX parsing, report generation, structure
  session/      Redis cache, context compression
  api/          FastAPI routes, job runner, middleware, schemas
  cli.py        CLI interface