AI-Powered Home Loan Guardian for Mortgage Refinancing

lendi · Finance · 2026 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Finance

Lendi wanted to make mortgage refinancing more proactive and personalized while preserving the trust of human broker relationships.

Constraints:

• financial conversations must stay within regulatory and communication boundaries
• customer context needs to persist across sessions
• agents need access to internal systems and external services
• consumer demand requires scalable deployment

Architecture

Lendi Guardian is a regulated-finance multi-agent application on Amazon Bedrock. Customers enter through a chat-led dashboard UI, API Gateway routes requests to AI agents running on EKS, specialized Bedrock agents collect loan information and recommend products, MCP servers connect internal and external systems, Bedrock Guardrails enforce mortgage communication boundaries, Langfuse and CloudWatch provide traces/logs, and MongoDB/S3 hold customer state and documents.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  guardian-chat-ui["guardian-chat-ui"]
  api-gateway-layer["api-gateway-layer"]
  eks-agent-runtime["eks-agent-runtime"]
  bedrock-agent-layer["bedrock-agent-layer"]
  mcp-integration-servers["mcp-integration-servers"]
  bedrock-guardrails["bedrock-guardrails"]
  customer-context-store["customer-context-store"]
  s3-document-store["s3-document-store"]
  langfuse-observability["langfuse-observability"]
  cloudwatch-logs["cloudwatch-logs"]
  broker-human-handoff["broker-human-handoff"]
  guardian-chat-ui --> api-gateway-layer
  api-gateway-layer --> eks-agent-runtime
  eks-agent-runtime --> bedrock-agent-layer
  eks-agent-runtime --> langfuse-observability
  eks-agent-runtime --> cloudwatch-logs
  bedrock-agent-layer --> mcp-integration-servers
  bedrock-agent-layer --> bedrock-guardrails
  bedrock-agent-layer --> customer-context-store
  customer-context-store --> s3-document-store
  bedrock-agent-layer --> broker-human-handoff

Components

Data flow

Guardian continuously monitors home loans and property context, alerts customers when better refinance options appear, guides the customer through a chat flow, hands context across specialized agents for engagement, information collection, lender/product matching, product-specific application preparation, and Linda re-engagement, then prepares applications while preserving broker involvement for high-value relationship work.

Why this architecture

Lendi needed a regulated, personalized, always-on refinance journey rather than a generic chatbot. Multi-agent separation maps to mortgage workflow stages, EKS supplies scalable agent hosting, Bedrock gives model choice and guardrails, MCP servers connect institutional systems, and observability is essential because financial decisions require traceability and continuous optimization.

Models used: Amazon Bedrock foundation models

Scale

• Guardian was built and deployed in 16 weeks.
  “Lendi used Amazon Bedrock to accelerate the build of their agentic solution within 16 weeks.”article, stated
• The launch followed a more than 30,000-hour cross-functional sprint.
  “Guardian launched in only 16 weeks following a more than 30,000-hour cross-functional sprint”article, stated
• Rate Radar scans thousands of home loans daily.
  “its AI-powered Rate Radar, which scans thousands of home loans daily”article, stated
• The refinance path is described as taking only 10 minutes with no paperwork or phone calls.
  “enables refinancing in only 10 minutes, with no paperwork, no phone calls, only a single tap.”article, stated
• Guardian had settled millions in home loans at faster refinance cycle times than Lendi's baseline.
  “already delivering measurable results, having settled millions in home loans with refinance cycle times considerably faster than Lendi Group’s baseline.”article, stated

Alternatives considered (roads not taken)

Treat AI only as an efficiency tool

Rejected because: Lendi framed Guardian as a reinvention of the home-loan journey that anticipates customer needs and preserves human relationship work.

Single monolithic refinance assistant

Rejected because: The refinance journey was split into specialized agents so concerns, maintenance, latency, and troubleshooting stayed manageable.

Use the most sophisticated model for every task

Rejected because: The lessons explicitly recommend task-appropriate model choice and domain knowledge instead of assuming the most sophisticated model is best.

Fully replace brokers with automation

Rejected because: The system is designed to free brokers for negotiation, empathy, and complex structuring.

Patterns

multi-agent-orchestration (orchestration)

The refinance flow is implemented as multiple specialized agents with distinct responsibilities.

tool-use-orchestration (integration)

MCP servers connect agents to institutional knowledge, internal systems, external data, and third-party services.

safety-policy-enforcement (governance)

Bedrock Guardrails constrain customer communications to regulatory and mortgage-relevant boundaries.

Tradeoffs

Use AI to handle routine guidance while preserving human broker relationship work.

Chose agentic support that frees human experts over treating AI only as an efficiency tool or replacing the human relationship.

Rationale: The source frames the goal as proactive guidance plus freeing human experts to focus on relationships.

Evaluation

No grounded evaluation method beyond observability traces is described in the available input.

Metrics: agent traces · performance metrics

Outcomes

• The ZenML summary reports development in 16 weeks and settled millions in home loans.

The source describes an architecture intended to scale on EKS for consumer demand.