Avijeet7/loan-underwriting-agent

======================================================================== AEGIS: AUTONOMOUS LOAN UNDERWRITING & AUDITING MULTI-AGENT PIPELINE ========================================================================

Live Link - https://avijeet7-loan-underwriting-agent.hf.space/

Aegis is an enterprise-grade, event-driven multi-agent orchestration pipeline designed to automate risk profiling, financial sanity auditing, and compliance monitoring for credit applications.

Powered by LangGraph and Gemini 2.5 Flash, the system decouples complex underwriting criteria into specialized, sovereign AI agents that collaborate via an asynchronous state-machine topology to eliminate human bias and accelerate credit decisioning pipeline latency.

1. 1.ARCHITECTURAL TOPOLOGY & AGENT ROLES ------------------------------------------------------------------------

The core engine is built as an explicit directed acyclic state graph (DAG) where context, short-term conversational history, and structural payloads are seamlessly synchronized across execution boundaries:

• —DATA INGESTION ENGINE: Validates incoming raw JSON profile streams, seeds state vectors, and evaluates presence of compulsory documentation.
• —FINANCIAL ANALYST AGENT: Evaluates balance sheets, operating vintage, Debt Service Coverage Ratio (DSCR), and CIBIL bureau scores.
• —COMPLIANCE AUDITOR AGENT: Audits the evaluation data against strict jurisdictional guidelines, tax verified status (PAN/ITR), and legal boundaries.
• —PYDANTIC SANITY GUARDRAIL: A compilation abstraction layer that intercepts final structured responses to guarantee 100% data schema compliance before serialization to the user interface.

1. 1.SYSTEM DIRECTORY TREE ------------------------------------------------------------------------

Autonomous-Loan-Underwriting-Auditing-Pipeline/ │ ├── app/ │ ├── _init_.py │ ├── main.py # FastAPI Application Server & Global Context │ ├── graph.py # LangGraph Topology & Specialized Nodes │ ├── schemas.py # Pydantic Enforcement Model Layer │ └── agents.py # Core Agent Prompts & LLM Tool Interfaces │ ├── frontend/ │ └── app.py # Streamlit User Interface & Real-time Ledger Stream │ ├── .env # Local Environment Variables (Git-Ignored) └── requirements.txt # Unified Project Software Dependencies

1. 1.PREREQUISITES & ENVIRONMENT SETUP ------------------------------------------------------------------------

Ensure you have Python 3.11+ installed on your local workstation environment.

Step 1: Clone the Project Repository $ cd Autonomous-Loan-Underwriting-Auditing-Pipeline

Step 2: Install Project System Dependencies $ pip install -r requirements.txt

*Note: If encountering interpreter environment compilation mismatches, run: $ python -m pip install -r requirements.txt

Step 3: Setup Google AI Studio API Access Key Create a .env file in the project's root directory:

GOOGLEAPIKEY=AIzaSy...[Your-Actual-Gemini-API-Key]

1. 1.RUNNING THE NATIVE DEVELOPMENT PIPELINE ------------------------------------------------------------------------

The framework utilizes a decoupled client-server architecture. You must initialize the asynchronous backend process before executing the frontend UI.

Step 1: Boot the FastAPI Asynchronous Server Instance Open a terminal window in the project root directory and run: $ uvicorn app.main:app --reload --port 8000

Ensure the console stream concludes with: "INFO: Application startup complete."

Step 2: Boot the Streamlit Operational Control Console Open a secondary distinct terminal session in the project root and run: $ streamlit run frontend/app.py

The frontend web platform will automatically bind and initialize inside your native default browser engine at: http://localhost:8501

1. 1.DOCKER CONTAINERIZATION & PRODUCTION DEPLOYMENT ------------------------------------------------------------------------

To preserve total infrastructure parity across staging, sandbox, and production cloud nodes, this architecture is fully Docker containerized.

To compile and execute the system container locally: $ docker build -t aegis-underwriting-pipeline . $ docker run -p 8501:8501 --env-file .env aegis-underwriting-pipeline

Production Destination (Portfolio Demonstration Hub):

• —This pipeline is tailored for immediate deployment onto Hugging Face Spaces using the Native Docker SDK container registry runner.
• —Set your Google AI Studio Credentials safely using Hugging Face's native Repository Secrets Management portal (GOOGLE_API_KEY) to keep deployment keys private.

1. 1.RUNTIME CONSTRAINTS & RATE LIMIT HANDLING (429 QUOTA) ------------------------------------------------------------------------

• —The Gemini 2.5 Flash Free Tier includes a strict baseline threshold of 20 Requests Per Day (RPD) per Project Container.
• —Because a singular autonomous profile evaluation triggers consecutive agent hops (Data Ingestion -> Financial Analyst -> Compliance Auditor), submitting 3-4 evaluation profiles sequentially can exhaust daily quotas.
• —Solution for extended demos: If a hard quota limits testing, log into Google AI Studio, instantiate a "New Project Workspace," pull a fresh API string, and drop it into your local workspace. ========================================================================