Project Memo
Project Memo
Using GenAI to cut mortgage document preparation from 2 days to 5 minutes.
Problem
A 200-page manual process
For every mortgage application in Hong Kong, a loan officer has to compile a project memo - a detailed briefing document that informs credit decisions. Creating one meant manually downloading up to 200 pages of supporting documents from a legacy system, opening each one, and filling in a template by hand. Complex nested tables, mixed English and Chinese content, and low-quality scans made it slow and error-prone. On average, it took 2 days. There was no digital support, no conditional logic, no extraction tools - just documents, a template, and time.
My role
Sole designer from discovery through handover, covering research, journey mapping, wireframes, prototyping, and iteration.
Team
Product owner, co-product owner (regulatory), data scientists, and engineers.
Process
Research first, solutioning second
3 things shaped how Project Memo came together, and each one tested a different aspect of the design process.
Built on the ground
We started with remote interviews - watching mortgage officers walk through their process and narrate their pain points in real time.
Creative constraints
The ideal solution was a split-screen interface: extracted results alongside the source document, with the relevant chunk highlighted. Engineering ruled it out. Instead, we worked with what Office 365 could do - hyperlinks embedded in the exported Word document that jump to the exact source page in the PDF. Not the original vision, but genuinely useful.
Logic before screens
Before any interface was designed, we mapped the conditional logic: which fields appear for which property types, what each field requires, and how the GenAI should extract and surface it. That groundwork is what made the product feel considered rather than generic.
The flow, simplified
Login, create a new case, upload documents, review the AI-extracted table, edit where needed, export to Word - complete with hyperlinks back to the source pages. From upload to download: under 5 minutes.
Outcome
2 days of document work, down to 5 minutes
Project Memo shipped and is actively in use by the mortgage team in Hong Kong. What once required up to 2 days of manual document review now takes minutes - users upload their supporting documents, review the AI-extracted fields, make any edits, and export a completed memo with embedded source references. The product has since been handed over to the retail banking UX team to continue building on.
Reflection
What GenAI product work teaches you that chatbot design doesn't
In-person research pays compound interest
Remote interviews got us the what. The Hong Kong trip gave us the why - and the relationship. Meeting users face to face changed how freely they talked, and the conversations we had over lunch unlocked context that hours of remote calls had missed.
Advocacy isn't just about the ideal
We pushed hard for the full document viewer solution. When engineering ruled it out, we didn't accept a lesser experience - we found a different path to the same goal. The hyperlink solution wasn't the plan, but it delivered what users actually needed.
Know when to ship the 80%
Working with data scientists added rigour we genuinely needed. But rigour can become a bottleneck. The unlock was deciding early: if it covers 80% of cases well, ship it. Edge cases can follow. Waiting for perfect delays the learning that only real use can give you.