System Architecture & Flow Map

Description: A systems-level backend architecture blueprint mapping how user inputs are safely classified, scrubbed for sensitive data, and dynamically routed between a deterministic API engine and a non-deterministic Large Language Model (LLM).
What I Did:
Designed a Dual-Path Hybrid Architecture: Recognizing that financial transactions require strict, hardcoded security while account inquiries benefit from natural language, I architected a system that forks user inputs. Generative AI handles conversational context, while money-movement tasks are strictly ring-fenced inside compliant API paths.
Engineered Multi-Layered Guardrails: Designed an asynchronous Input Guardrail Layer that intercepts customer inputs to redact or mask PII (Personally Identifiable Information) before it ever hits an external LLM wrapper.
Created Post-Processing Verification: Built an Output Guardrail Layer checking for system hallucinations, brand alignment, and accuracy before rendering the final text back to the customer, ensuring 100% adherence to strict financial regulatory standards.
Strategic Impact:
This blueprint solved the enterprise dilemma of leveraging generative AI safely in a highly regulated banking environment. It prevented data leaks, maintained compliance, and lowered conversational friction.