🔬 AI Research Paper Reviewer

An AI system that reads a research paper and writes a structured peer review — then grades its own work and gets better over time.

---

What it does

You give it a PDF  →  It writes a review  →  It checks the review  →  It learns from mistakes

Step by step:

1. Reads the paper — extracts title, abstract, methods, experiments, claims, and contributions
2. Finds related work — searches Semantic Scholar for similar papers
3. Checks novelty — compares the paper's claims against prior work
4. Writes a review — generates a structured NeurIPS/ICML-style review with scores
5. Grades itself — a second AI pass checks the review for hallucinations, missing content, and unsupported claims
6. Improves — saves quality scores and automatically adjusts settings on the next run

---

Quickstart

# 1. Install dependencies
pip install -r requirements.txt

# 2. Add your API key
cp .env.example .env
# Edit .env and set: ANTHROPIC_API_KEY=sk-ant-...

# 3. Run the web app
streamlit run app.py

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

---

Web Interface

Upload a PDF → choose a venue → click ▶ Run Review

Tab	What you see
📄 Review	The full peer review with score, strengths, weaknesses, and questions
🧑‍⚖️ Evaluation	Quality scores — was the review grounded? did it hallucinate?
📚 Related Papers	Papers found on Semantic Scholar + novelty assessment
📊 Metrics Dashboard	History of all past runs and performance trends

---

CLI (optional)

# Basic review
python main.py paper.pdf

# Specify venue
python main.py paper.pdf --venue ICML

# Add custom focus
python main.py paper.pdf --criteria "Focus on fairness and reproducibility"

# Save full output as JSON
python main.py paper.pdf --output review.json

# View performance history
python main.py --stats

---

How it self-improves

After each review, quality scores are saved locally. Before the next run, the system reads those scores and adjusts automatically:

Problem detected	What changes
Review makes things up	Switches to a stricter prompt
Review misses paper content	Fetches more related papers, uses a more thorough prompt
Review quality is low	Runs a second self-critique pass

---

Supported venues

NeurIPS · ICML · ICLR · ACL · CVPR · Generic

---

Project structure

├── app.py              ← Streamlit web interface
├── main.py             ← CLI interface
├── src/
│   ├── extractor.py    ← Reads PDF, extracts paper structure
│   ├── retrieval.py    ← Searches Semantic Scholar, checks novelty
│   ├── reviewer.py     ← Generates the review
│   ├── evaluator.py    ← Grades the review quality
│   ├── monitor.py      ← Saves and displays metrics history
│   ├── adaptive.py     ← Adjusts settings based on past performance
│   └── schemas.py      ← Data models
├── data/metrics/       ← Review history (auto-created)
└── requirements.txt

---

Requirements

• Python 3.10+
• Anthropic API key (get one here)
• Optional: Semantic Scholar API key for higher rate limits