How the ChatBot Helped Vine Financial’s Customers 

• Time lost searching for deal information: Lending teams had to manually cross-reference loan data in one system with details buried in tax returns, credit memos, and appraisals in another. The chatbot now answers a single question by pulling from both sources at once, cutting the time spent hunting for information.
  
• Difficulty getting a complete picture of a deal: Questions like whether a loan could be supported by cash flow required piecing together loan terms, financials, and collateral data from multiple places, often leaving gaps. The chatbot now plans out everything a question requires before answering, so lending teams get a complete response the first time.
  
• Slow document review on large loan files: A single loan package can include dozens of documents, making it slow and tedious for underwriters to locate the one relevant clause or figure. The chatbot searches across all of a deal's documents in seconds and surfaces the exact passage that answers the question.
  
• Risk of costly errors in underwriting: Manually calculating or transcribing figures like DSCR and LTV carries the risk of mistakes that can lead to poor credit decisions. The chatbot only reports verified figures, never estimates a value it cannot confirm, and shows the source behind every number.
  
• Confidence in data privacy across organizations: Lending organizations sharing a platform needed assurance that their deal data would never be visible to another organization. The chatbot enforces strict data separation, so every answer is limited to the requesting organization's own deals.
  
• Expectation of a fast, modern experience: Lending teams expect the same speed and ease they get from everyday digital tools. The chatbot delivers a complete, sourced answer within a few seconds, even for complex questions.

Why AWS?

• AWS Bedrock provided secure, enterprise-grade access to Claude models without the overhead of managing dedicated infrastructure, and allowed the same model family to power multiple stages of the pipeline: intent classification, query planning, and response synthesis.
• Amazon OpenSearch and Amazon Titan Embeddings gave the project a proven, scalable foundation for semantic document search, integrated natively with the rest of the AWS environment.
• AWS Cognito and Secrets Manager supported the strict multi-tenant security and credential management required for a financial services application.
• Amazon ECS, S3, and CloudFront delivered a managed, scalable infrastructure layer, allowing the engineering team to focus on the AI pipeline rather than server management.
• The breadth of AWS's managed AI and data services let Tech Holding assemble a compliant, production-ready architecture on a single cloud platform, rather than integrating multiple vendors.

How Tech Holding Built the Chatbot

• Intent classification and routing: Every question is classified as conversational, recall, quantitative, qualitative, or hybrid, and routed to the appropriate data source - database, documents, or both in parallel.
• Query decomposition: Before retrieval begins, an AI planning step breaks complex questions into a structured plan identifying exactly which data types and document searches are needed, and builds a completeness checklist so the final answer cannot omit key information.
• Structured data retrieval: Relevant facts (loan terms, spreading, collateral, existing loans) are pulled from the database in parallel queries, with security rules that always scope results to the requesting organization.
• Document search: Uploaded documents are broken into searchable chunks and embedded as vectors, enabling semantic search that finds the most relevant excerpts even across large loan packages.
• Response synthesis with guardrails: A final AI step combines database results and document excerpts into a single answer, governed by strict rules: only use retrieved data, never calculate a ratio when inputs are missing, and cite a source for every factual claim.
• Domain knowledge layer: A shared knowledge module defines lending terminology, deal parties, and data relationships, ensuring every stage of the pipeline interprets financial questions consistently.

Results and Benefits for Vine Financial

• Accurate, cited financial answers: Lending teams receive serviceability, DSCR, LTV, and policy answers grounded in actual deal data and documents, with anti-hallucination controls preventing fabricated figures.
• Unified access to deal information: A single natural-language question now returns an answer that blends structured loan data with document content, eliminating the need to search multiple systems.
• Seamless platform integration: The assistant lives directly inside the Vine platform as a deal-scoped sidebar, with conversation history retained for follow-up questions.
• Enterprise-grade compliance: Organization-level data isolation, audit logging, and citation-backed responses align the assistant with the compliance expectations of commercial lending.
• A scalable foundation for growth: The microservices architecture, built on AWS Bedrock, OpenSearch, and MySQL, supports future expansion of the assistant's capabilities as Vine's platform grows.

AWS Services Used