Ratnadeep Kakade

CICERO is designed to address the challenges of analyzing large, complex legal documents, researching, and developing evidence-backed legal arguments. It enables users to create legal cases (similar to Projects in ChatGPT or Claude), upload PDF/DOCX filings (including documents of approximately 500–600 pages), and interact with their cases through natural-language queries. Technical Architecture: CICERO follows a distributed, 5-plane architecture comprising Presentation, Control, Ingestion & Retrieval, Inference, and Data planes. The frontend uses React, Vite, TypeScript, and Zustand, deployed on Vercel. The backend uses Python, FastAPI, and Modal for API orchestration and compute-intensive workloads. Google Cloud provides PostgreSQL with pgvector for structured data and vector embeddings, Cloud Storage for document storage, and Pub/Sub for asynchronous processing. Workload Identity Federation enables keyless authentication between Modal and Google Cloud. Structure-aware RAG: CICERO uses Docling for PDF/DOCX parsing and OCR, followed by legal structure extraction covering clauses, definitions, cross-references, dates, and tables. Hierarchical parent-child chunking preserves document context. BGE-M3 generates embeddings, while hybrid retrieval combines dense vector search, PostgreSQL full-text search, and exact legal-identifier matching. Embedding and reranking workloads run on a Modal T4 GPU. Reciprocal Rank Fusion (RRF), BGE reranking, and context expansion improve retrieval relevance. Multi-agent Reasoning: The system comprises 7 specialized agents: Request Interpreter, Central Orchestrator, Document Intelligence, Legal Researcher, Argument Agent, Counterargument Agent, and Evaluator. Muse Spark-1.3 powers the Interpreter, Orchestrator, and Document Intelligence agent. DeepSeek-R1-Distill-Llama-70B (AWQ 4-bit), served through vLLM on a Modal A100 80GB GPU, powers the four specialist reasoning agents. A YAML-based Platform Capability Registry, validated using Pydantic, provides the Interpreter and Orchestrator with structured information about available agents, tools, capabilities, permissions, and workflows. The Interpreter converts user requests into structured task specifications, while the Orchestrator generates and coordinates execution plans. Gemini's 'Grounding with Google Search' supports external legal research. A Python execution harness manages task dependencies, parallel execution, state transitions, retries, permissions, and validation through structured agent contracts and persisted results. Citation verification and evidence validation prevent unsupported document claims from entering the final response. If sufficient evidence cannot be retrieved, the system explicitly reports insufficient evidence instead of fabricating conclusions.