Financial Underwriting Automation
A mid-market financial services firm was drowning in unstructured PDF documentation. Underwriters were spending hours manually extracting data across dozens of unstandardized forms to make a single risk assessment.
Before
4 Hours
per assessment
After Visthar
8 Seconds
per assessment
The Architecture
100% Data Privacy
Deployed a completely local, self-hosted LLM (Llama 3 70B via vLLM) on their private cloud. Zero data left their network.
Multi-Stage RAG Pipeline
Engineered a sophisticated Retrieval-Augmented Generation pipeline using Qdrant vector databases to accurately parse unstructured financial PDFs.
Deterministic Output Enforcement
LLMs hallucinate. We built a strict validation layer that forces the model to output exact JSON schemas, directly feeding into their existing legacy CRM.
Tech Stack