Fintech · AI Pipeline2026
MortgageAssistPro
Built the AI document extraction pipeline that processes UK mortgage ESIS and offer letters — designed a 3-tier fallback ladder (raw text → regex → LLM) to avoid unnecessary AI calls, then cut LLM token costs by 80% with prompt caching layers.
- Role
- Full Stack Engineer — AI extraction ladder, prompt caching, metered billing
- Timeline
- 2026 · Live
- Stack
- Node.jsTypeScriptAWS LambdaOpenAIPrompt CachingReactPostgreSQL
01
Screens & Architecture
02
The Problem
Processing UK mortgage ESIS and offer letters manually was slow and costly, while high LLM token consumption created significant API overhead.
03
How I built it
- Built a 3-tier extraction ladder (rawText template match → regex parsing → LLM agent) to avoid unnecessary AI calls on simple documents.
- Implemented prompt caching layers on AWS Lambda, cutting LLM token costs by 80% while improving processing speed.
- Architected a unified DRY metered subscription handler shared across the Offer Check Pro and C2C modules.
- Built the admin budget planner and annualised variable pay UI with optimized classification logic.
04
Outcome
- 80% reduction in LLM token costs through multi-tier extraction fallback and prompt caching.
- Zero-downtime metered billing system handling subscription resets across OCP and C2C.