- AWS Bedrock
- Python
- ML lead scoring
- LLM briefing generation
Sole architect on a ten track lending and mortgage accelerator
A US lending and mortgage AI accelerator spanning ten use cases, three in production and seven on the roadmap, at roughly a two week build cadence per track. Shipped tracks cover document automation, lead scoring, and lock-in extension recommendations.
Problem
Ten separate lending and mortgage use cases, each of which could plausibly be its own project. Built that way, the tenth track costs as much as the first and nothing learned on the way carries forward. The target was roughly a two week build cadence per track, which rules out starting from scratch each time.
Approach
I architected all ten tracks, three of which are in production and seven of which sit on the roadmap. The first is AWS Bedrock document automation, covering mortgage and lending document extraction across all standard document types, with cross document verification and external validation checks. The second is smart lead scoring: behavioural signals captured from the client’s website feed an ML scoring model, and an LLM writes the loan officer a briefing that marks the lead hot or cold, supplies talking points, and drafts the outreach email. The third is a lock-in extension recommendation agent that aggregates credit score deltas, new loans, and profile changes against historical comparable decisions, then produces a recommendation brief.
Architecture
The three shipped tracks sit on a shared Bedrock and data layer and deliver into the same loan officer workflow, so each new track inherits the plumbing rather than rebuilding it.
The hard part
The interesting engineering claim here is not any single track, it is repeatability. Ten tracks at a two week cadence only works if the extraction, data, and delivery layers are shared, and shared layers only stay shared if one person is holding the architecture across all of them. That is what being sole architect across the whole accelerator bought.
Outcome
Three tracks in production, seven architected and on the roadmap, at roughly two weeks per track. Every shipped track lands in the loan officer’s existing workflow rather than in a separate tool.
Next