Quality assurance · Production

Rexera uses LangGraph to reduce QC false positives to 2% in real estate transaction workflows

The problem

Rexera's QC application processed thousands of real estate workflows daily, but initial single-prompt LLM checks lacked the context and multi-step reasoning needed to handle complex, multi-dimensional real estate scenarios, producing high rates of incorrect issue flags.

First attempt

Single-prompt LLMs produced false positive rates of 35% and false negative rates of 10%. A multi-agent CrewAI approach improved both figures but agents still sometimes took incorrect decision paths in complex scenarios, leaving residual false positive and false negative rates.

Workflow diagram · grounded in source
1
Workflow enters QC application
trigger
“Rexera has developed a specialized QC application that reviews thousands of workflows daily”
2
Rush vs standard routing
routing
“When the application identifies a rush order, it's directed down the "Rush Order" branch of the tree. For standard orders, the application follows a different branch, focusing on regular processing checks”
3
LangGraph decision tree traversal
ai_action
“LangGraph, Rexera created a tree-like structure for the Quality Control (QC) application that allows for cycles and branching. This structure enables the QC application to navigate different paths based on the rush requirements”
4
Multi-dimensional QC checks
validation
“checks for errors across various stages of a real estate transaction, including data handling, client communication, and interactions with counterparties such as homeowner associations (HOAs), county offices, utility companies, and others”
5
Human-in-the-loop review
human_review
“the integration of human-in-the-loop workflows”
6
Issue flagging output
output
“proactively identifying issues to prevent delays”
Reported outcome

LangGraph's tree-like decision structure reduced false positives from 8% to 2% and false negatives from 5% to 2%, delivering greater consistency and accuracy across real estate transaction workflows.

Reported metrics
false positive rate — single-prompt LLM baseline35%
false positive rate — after CrewAI8%
false negative rate — single-prompt LLM baseline10%
false negative rate — after CrewAI5%
Show all 6 reported metrics
false positive rate — single-prompt LLM baseline35%
false positive rate — after CrewAI8%
false negative rate — single-prompt LLM baseline10%
false negative rate — after CrewAI5%
false positive rate — after LangGraph2%
false negative rate — after LangGraph2%
Reported stack
LangChainLangGraphLLMsCrewAI
Source
https://blog.langchain.dev/customers-rexera/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

LangGraph's tree-like decision structure reduced false positives from 8% to 2% and false negatives from 5% to 2%, delivering greater consistency and accuracy across real estate transaction workflows.

What tools did this team use?

LangChain, LangGraph, LLMs, CrewAI.

What results were reported?

false positive rate — single-prompt LLM baseline: 35%; false positive rate — after CrewAI: 8%; false negative rate — single-prompt LLM baseline: 10%; false negative rate — after CrewAI: 5% (source-reported, not independently verified).

What failed first in this deployment?

Single-prompt LLMs produced false positive rates of 35% and false negative rates of 10%.

How is this quality assurance AI workflow structured?

Workflow enters QC application → Rush vs standard routing → LangGraph decision tree traversal → Multi-dimensional QC checks → Human-in-the-loop review → Issue flagging output.

Related quality assurance cases