Zainab4626/auto-legal-analyst

📄 Contract Analysis AI

Fine-tuned LLM pipeline that extracts clauses, identifies legal entities, and performs risk analysis on multi-hundred-page contracts — cutting review time by 84% for autonomous vehicle firms.

---

🚀 Key Results

Metric	Result
Analysis Time Reduction	40% faster with fine-tuned LLaMA-7B
Contract Parsing	500-page contracts parsed in 2 minutes
Legal NER F1-Score	87% (+23 points over base model)
Review Time Savings	84% reduction for AV firm legal teams

---

🏗️ Architecture

The pipeline follows a 5-stage extraction and analysis workflow:

1. PDF Ingestion — Multi-format parsing via pdfplumber, PyMuPDF, with Tesseract OCR fallback
2. Chunking & Preprocessing — Recursive 512-token splitting with 50-token overlap using LangChain
3. Fine-tuned LLM Inference — LoRA-adapted LLaMA-7B (rank-16, 4-bit quantized) for clause classification
4. NER & Clause Extraction — Custom legal NER identifying parties, dates, obligations, liability clauses
5. Risk Analysis & Output — Structured JSON with risk scores, flagged clauses, and summary reports

---

🧠 Model Details

Fine-Tuning Configuration

from peft import LoraConfig, TaskType

lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"]
)

• Base Model: LLaMA-2-7B (meta-llama/Llama-2-7b-hf)
• Training Data: 5,000+ annotated legal documents
• Quantization: 4-bit via bitsandbytes for efficient GPU inference
• Adapter: LoRA rank-16 on all attention projection layers

Performance vs Base Model

Task	Base LLaMA-7B	Fine-tuned (LoRA)	Improvement
Legal Entity Recognition	64% F1	87% F1	+23 pts
Clause Classification	58% Acc	91% Acc	+33 pts
Obligation Detection	52% F1	83% F1	+31 pts
Risk Flag Precision	45% Prec	79% Prec	+34 pts

---

🎯 Interactive Demo

The project includes a live demo where you can paste contract text and see clause extraction with risk analysis in real-time.

Features:

• Paste any contract text or load a sample
• Clause type classification (Indemnification, Liability Cap, etc.)
• Risk scoring (High / Medium / Low)
• Named entity extraction (parties, dates, monetary values)
• Processing time benchmarks

---

🔧 Tech Stack

Category	Technologies
LLM	LLaMA-2-7B, LoRA, PEFT, bitsandbytes
ML Framework	PyTorch, HuggingFace Transformers
NLP Pipeline	LangChain, spaCy, tiktoken
PDF Processing	pdfplumber, PyMuPDF, Tesseract OCR
Data Validation	Pydantic
API	FastAPI
Frontend	React, TypeScript, Tailwind CSS, Framer Motion

---

📁 Project Structure

├── model/
│   ├── fine_tune.py          # LoRA fine-tuning script
│   ├── config.yaml           # Training hyperparameters
│   └── evaluate.py           # Benchmark evaluation
├── pipeline/
│   ├── extract_clauses.py    # Clause extraction pipeline
│   ├── ner_model.py          # Legal NER module
│   ├── risk_scorer.py        # Risk analysis engine
│   └── pdf_loader.py         # PDF ingestion utilities
├── api/
│   └── main.py               # FastAPI endpoints
├── data/
│   ├── train/                # Training dataset (5,000+ docs)
│   └── eval/                 # Evaluation benchmark
└── frontend/                 # React demo application

---

🏢 Business Impact

Client: Autonomous vehicle manufacturing firms

• Reduced legal contract review time from ~6 hours to under 1 hour per contract
• Enabled legal teams to process 5x more contracts per quarter
• Automated flagging of high-risk clauses (indemnification, liability caps)
• Extracted and structured key entities across 500-page procurement agreements

---

📬 Contact

Built as part of my AI Engineering portfolio. Open to discussing the technical implementation, model architecture decisions, or potential applications.

---

_{Built with PyTorch · HuggingFace · LangChain · LoRA}