How to build a Validation Agent

TL;DR

This agent automates the validation of borrower real estate track records by cross-referencing uploaded documents, SharePoint records, and public web data, then generates a structured, auditable JSON report for underwriters.

What It Does:

• Collects borrower name/ID and uploads their track record files (spreadsheets, PDFs).

• Searches SharePoint for historical loan and ownership data.

• Searches the web for county PVA, warranty deeds, and HUD statements.

• Uses an AI agent to parse, cross-check, and validate all data sources.

• Flags discrepancies, missing data, or compliance risks.

• Outputs a structured JSON validation report.

Who It’s For:

• Real estate loan underwriters

• Lending operations teams

• Compliance and risk analysts

Time to Value:

• Immediate: Upload files and enter borrower info, get a validation report in minutes.

• No manual cross-checking or data entry required.

Output:

• A structured JSON validation report containing:

  • Borrower Name / ID

  • Verified Loan Count

  • Discrepancies flagged (if any)

  • Compliance notes (HUD/ownership validation)

  • Confidence score

Common Pain Points for Validation

• Manual, time-consuming cross-checking of multiple data sources

• Risk of missing discrepancies or compliance issues

• Inconsistent documentation formats (PDFs, spreadsheets, scanned docs)

• Difficulty accessing and searching internal (SharePoint) and external (web) records

• Lack of audit trail or structured output for compliance

What This Agent Delivers

• Automated extraction and validation of borrower track records

• Cross-referencing across uploaded files, SharePoint, and public web data

• Discrepancy and compliance risk flagging

• Structured, auditable JSON output

• Drastically reduced manual effort and turnaround time

Step-by-Step Build (StackAI Nodes)

1) Borrower Name/ID (Input Node)

What it does:

• Collects the borrower’s name or ID for use in all downstream searches and validation.

Goal:

• Ensure all data sources are queried for the correct borrower.

2) Borrower Track Record Files (Files Node)

What it does:

• Lets users upload borrower track record files (spreadsheets, PDFs).

• Extracts and processes text, including OCR for scanned documents.

Goal:

• Provide the AI with the borrower’s historical real estate activity.

3) SharePoint Search (Action Node)

What it does:

• Searches SharePoint for historical loan and ownership data using the borrower’s name/ID.

Goal:

• Retrieve internal records for cross-validation.

4) Web Search (Action Node)

What it does:

• Searches the web for county PVA, warranty deeds, and HUD statements related to the borrower.

Goal:

• Gather public records for additional validation.

5) AI Underwriting Verification Agent (LLM Node)

What it does:

• Parses all collected data.

• Cross-checks details across sources.

• Flags discrepancies, missing data, or compliance risks.

• Generates a structured JSON validation report.

Goal:

• Automate and standardize the underwriting validation process.

Instructions

You are an AI Underwriting Verification Agent. Your role is to verify borrower real estate track records as part of loan underwriting. You must extract, cross-check, and summarize details into a structured validation report. Always flag discrepancies, missing data, or potential compliance risks. Be concise and accurate.



Inputs:

- Uploaded borrower track records (spreadsheets, PDFs).

- County PVA websites, warranty deeds, HUD statements (via linked knowledge base/web search).

- Historical loan and ownership data stored in SharePoint.



Workflow Steps:

1. Parse the borrower’s track record → extract purchase/sale dates, amounts, ownership details, and loan history.

2. Cross-check these details against county PVA sites, warranty deeds, and HUD statements.

3. Identify discrepancies or gaps → mark them as “Exceptions for Review.”

4. Generate a structured validation report that includes:

  • Borrower Name / ID

  • Verified Loan Count

  • Discrepancies flagged (if any)

  • Compliance notes (HUD/ownership validation)

  • Confidence score



Output Format:

Return a JSON object with the following fields:

- borrower_name

- verified_loan_count

- discrepancies

- compliance_notes

- confidence_score

Prompt

Borrower Name/ID: {{in-0}}



Borrower Track Record:

{{doc-0}}



SharePoint Data:

{{action-0}}



Web Search Results:

{{action-1}}



Generate a JSON validation report as described in the system prompt

6) Output (Output Node)

What it does:

• Displays the final JSON validation report to the user.

Goal:

• Provide a clear, auditable result for underwriters and compliance teams.

7) Add Column (Action Node)

What it does:

• Adds a new column to a Smartsheet for tracking or logging results (optional, for workflow integration).

Goal:

• Enable easy tracking and further automation in Smartsheet.