cuad

AI-powered contract review skill with CUAD risk detection, market benchmarks, and lawyer-ready redlines. Works with Claude Code, Codex, Cursor, and 26+ tools.

Benchmark Microsoft Foundry Content Understanding on CUAD legal contracts. Achieves 83.3% F1 score (29% better than GPT-4o baseline). Complete Python notebook with optimized schemas for contract clause extraction. Production-ready with confidence scores & cost analysis.

Document Q&A with verifiable citations over CUAD contracts: BM25 retrieval, local LLM via Ollama, grounded quotes, abstention and evaluation

Pre-embedding definition injection for legal contract RAG — fixes the Definition Dependency Gap in CUAD contracts. Code and benchmark for the DAPEI paper.

Check contracts for clauses on an ordinary PC with a small local model (Jeff). Measured on 102 lawyer-labelled contracts.

Benchmark zero-shot LLM vs fine-tuned transformers for contract clause classification on the CUAD dataset: precision, recall, F1, cost & latency side by side.

Evaluation of Llama 3.2 3B on CUAD Dataset - before and after finetuning

Multi-agent LLM-based contract review system — Summer School Turco 2026

End-to-end Agentic RAG system for legal contract analysis — hybrid retrieval (BM25 + dense + RRF), cross-encoder reranking, LangGraph agent with 3 tools, LLM-as-a-judge evaluation, FastAPI + Streamlit + Docker Compose. Built on CUAD (406 contracts, 22K expert annotations). Zero cost — all free-tier tools.

Agente de RAG contratual avaliado pelo que ele RECUSA responder — CUAD: 510 contratos, 20.910 perguntas, 68% sem resposta.

AI contract analysis: clause extraction, risk flagging, and cited Q&A — honestly evaluated against lawyer annotations (CUAD)

Professional commercial contract search platform powered by Qdrant vector search and the CUAD dataset. Features 510 real contracts with 13K+ expert-labeled clauses, semantic search, and coarse-to-fine retrieval pipelines.

AI-powered contract intelligence & risk engine with open-vocabulary clause classification, interactive systemic risk graphs, and real-time What-If redline simulations.

QLoRA fine-tuning of Qwen2.5-7B on CUAD legal contracts with a rigorous before/after eval: clause classification 61.5%→80%, missed clauses 57%→5%, calibration error 10× lower, hallucinations hand-audited (1.0%→3.5%). Trained on a free T4, runs on a 6 GB GPU, with a FastAPI + Streamlit demo.

Private contract analyst: an open blueprint for private document AI with open models on your own GPU (CUAD + ContractNLI evaluation, verified quotes). Not legal advice.

Ask a real contract, get the clause: RAG over 100 CUAD contracts with local open models, a 189-setup search ablation, claim-level grading and a failure taxonomy.

QLoRA fine-tune of Qwen3-4B on CUAD — 41 clause types, verbatim extraction, evaluated against zero-shot frontier models

Audits contract PDFs against rule playbooks with hybrid RAG (BM25 + dense + reranking); every finding cites verified quotes. Runs locally with Ollama or with OpenAI. Benchmarked on CUAD.

Fine-tuned Qwen3-4B for contract review: vLLM vs Ollama, FP8/AWQ/Q4 accuracy, load tests to 32 users, cost per 1,000 contracts, and a comparison with Claude Opus 5.5. Docker-deployable.