Land-Sentinel AI

A Multi-Flow AI pipeline with 5 agents that detects document fraud, verifies land records, and delivers real-time underwriting decisions for banking institutions.

Comment

Problem

Reviewing Land Documents Manually is very inefficient and detecting any fraud in those document are very hard. Due to this bank staff have to do a lot of manual work when it comes to review documents in order to provide Loans.

Proposed Solution

Land-Sentinel AI is a multi-agent document fraud detection and regulatory compliance pipeline for banking underwriters. It processes uploaded land records, Aadhaar cards, PAN cards, and legal documents through 5 specialized AI agents and produces a final underwriting decision.

Impact

  • Reduces the Manual Document review time.
  • Easy Detection of any tampering made in the documents.
  • Stay Consistent with data throughout the Review Process.
  • Cuts LLMs API cost by stopping other agents call when tampering is detected.
  • Summaries each agent output in an understandable format.

The 5-Agent Pipeline

Agent Role
🔍 Forensic Agent Scans documents for tampering, forged stamps, inconsistent fonts using Vision LLM
📋 Extraction Agent Pulls structured fields — borrower name, PAN, income, loan amount, property ID
🔗 Verification Agent Cross-checks extracted data against land registry, ITR records, and PAN database
⚖️ Regulatory Agent Evaluates 5 Measurable Action Points — DTI ratio, LTV, document age, income floor
📊 Synthesizer Agent Combines all findings into a final decision with risk score and audit trail

Final Decisions

  • APPROVED — all checks pass, compliant
  • REJECTED — document tampering detected
  • REVIEW — verification mismatch, needs human review
  • NON_COMPLIANT — regulatory rules failed

How We Built It

The core orchestration layer is built with LangGraph — a stateful multi-agent graph framework that handles conditional routing between agents. If the Forensic Agent detects tampering above a threshold score, the pipeline short-circuits immediately without calling downstream agents, saving cost and time.

LLM inference runs on Groq's LPU hardware using three models selected by task complexity:

  • llama-4-scout-17b for vision-based document analysis
  • llama-3.3-70b-versatile for structured extraction and regulatory reasoning
  • llama-3.1-8b-instant for lightweight verification logic and synthesis

The verification layer is architected with an abstraction interface designed to connect to real government-backed APIs Surepass Land Record API In the current prototype we have used Mock data.

The frontend is a Streamlit app with a ChatGPT-style session model — each document upload creates a new application session with its own pipeline run, stored in session history in the sidebar.

Built With

  • api
  • groq
  • langgraph
  • llama
  • pydantic
  • pypdf2
  • python
  • streamlit
  • surepass
Share this project:

Updates