Team Null Set Ayush Kumar Singh | Isha Singh | Priyanka Gnana Karanam Newton School Of Technology

---

---

Project Overview

ContraLegal AI is an autonomous legal intelligence platform engineered to transform unstructured contract data into actionable risk distributions. By synthesizing Legal-BERT transformer architectures with Retrieval-Augmented Generation (RAG), the system provides granular multi-class risk scoring, automated redrafting, and spatial PDF highlights to eliminate manual bottlenecks in enterprise legal review.

---

Core Engineering Capabilities

Capability	Orchestration	Technical Specification
Trimodal Classification	Legal-BERT	High, Medium, and Low-intensity risk granularity
Dynamic Scoring	Hybrid Logic	Fusion of Transformer probabilities and deterministic keyword heuristics
Explainable AI	RAG + LangChain	Root-cause analysis of flagged clauses in professional nomenclature
Strategic Redrafting	Generative AI	Automated generation of balanced, legally-sound alternative phrasing
Conversational Querying	FAISS Vector Store	Real-time, document-grounded Q&A for complex legal inquiries
Spatial Annotation	PyMuPDF API	Physical coordinate-to-text mapping for in-situ PDF highlighting
Relational Data Export	openpyxl	Structured synthesis of risk distributions in Excel and CSV formats
Thematic Clustering	Scikit-Learn	Unsupervised K-Means grouping of obligation-specific clauses

---

Engineering Hierarchy & Contributions

The platform is built upon a high-concurrency architecture with a strict separation of concerns across research and deployment layers.

Deep Learning & Transformation | Ayush Kumar Singh

• Fine-tuned the nlpaueb/legal-bert-base-uncased transformer using a weighted-trainer objective for imbalanced class distribution.
• Engineered the 3-class quantitative heuristic for synthetic label generation spanning over 21,000 samples.
• Developed the formal ablation study and multi-class ROC-AUC evaluation suite to validate transformer superiority over statistical baselines.

Generative AI & RAG Orchestration | Priyanka Gnana Karanam

• Architected the retrieval-augmented generation pipeline utilizing FAISS for vectorized similarity search.
• Engineered the LLM Provider Factory, enabling seamless interoperability between Google Gemini, Groq, and OpenAI.
• Validated prompt-engineering strategies for deterministic clause synthesis and document-grounded conversational flows.

Spatial NLP & Deployment Systems | Isha Singh

• Engineered the spatial highlighting engine using PyMuPDF to perform physical document marking via bounding-box coordinate tracking.
• Implemented semantic document segmentation to optimize transformer context windows.
• Architected the automated CI/CD infrastructure via GitHub Actions for continuous environment validation.

---

Quantitative Performance Matrix

The integration of transformer architectures resulted in a fundamental shift in both classification precision and recall intensity.

Metrical Indicator	Random Forest Baseline	Legal-BERT Transformer	Improvement (Δ)
Accuracy	94.44%	97.01%	+2.57%
Weighted F1	0.9441	0.9702	+2.76%
Macro F1	0.8901	0.9371	+5.28%
ROC-AUC (Macro)	0.9870	0.9948	+0.79%
High Risk Recall	73.96%	85.94%	+11.98%

Note: The 11.98% surge in High Risk Recall represents the most critical engineering milestone, ensuring safety in mission-critical legal review.

---

Global Repository Schema

ContraLegal-AI/
├── app.py                          # Streamlit Production Environment
├── .github/workflows/              # Automated CI/CD (python-app.yml)
├── src/
│   ├── model_trainer.py            # Phase-integrated Training Orchestrator
│   ├── data_pipeline/              # Semantic Extraction & Normalization
│   ├── inference/
│   │   ├── predictor.py            # Trimodal Detection Engine (BERT/RF)
│   │   ├── llm_engine.py           # RAG Orchestrator & Conversational Layer
│   │   └── keyword_engine.py       # Deterministic Rule Definitions
│   ├── model/
│   │   ├── bert_trainer.py         # Transformer Fine-tuning Suite
│   │   └── evaluator.py            # Quantitative Performance Metrics
│   └── utils/
│       └── pdf_annotator.py        # Spatial Coordinate Highlighting
├── models/
│   ├── legal_bert/                 # Fine-tuned Weights (nlpaueb)
│   └── ablation_study.png          # Baseline vs. Transformer Visualization
├── notebooks/
│   └── train_legal_bert_colab.py   # GPU-accelerated Training Script
└── report/
    ├── report.pdf                  # Formally Published IEEE Paper
    └── report.tex                  # Scientific Manuscript Source

---

Operational Deployment

Environment Initialization

git clone https://github.com/AyushCoder9/ContraLegal-AI.git
pip install -r requirements.txt

Application Execution

To initiate the production dashboard with the global pre-trained model:

streamlit run app.py

Analytical Training (Optional)

To execute the full analytical pipeline and regenerate performance artifacts:

python -m src.model_trainer

---

Scientific Publication

The technical methodology, algorithmic decisions, and empirical evaluations are documented in the associated IEEE conference-format manuscript located in the report/ directory.

---

Null Set | 2026
Engineered for Legal Precision.