Mortgage lending · Project 05
Two AI products for a document-heavy lender: a chat assistant grounded in every loan product and competitor product, and an agentic pipeline that turns unsorted bank statements into totals, insights and loan recommendations in 15–20 minutes.

- Domain
- Mortgage lending
- Build
- 4-month engagement
- Corpus
- Thousands of complex PDFs
- Statement review
- 2–3 days → 15–20 min
Project impact
Bank statement review cut from days to minutes.
The brief
Mortgage lending runs on documents. GreenBox asked for two products: one that could answer from everything the business knows, and one that could do the slowest document review in the building.
Mortgage eligibility, pricing, and policy answers lived across thousands of internal product documents. Finding the right one meant knowing it existed first, and the people who knew were the bottleneck.
Separately, reviewing a borrower's bank statements took two to three days of manual work. Statements arrived unsorted, as PDFs across multiple banks and date ranges, and someone had to read every page to classify credits against debits and work out the totals before a loan officer could recommend a product.
The deliverables
- 01
MIA, a custom RAG-supported chat assistant with full context of every GreenBox loan product and competitor product. We indexed and parsed thousands of complex PDFs so staff can ask in plain language and get an answer grounded in the right documents.
- 02
Two specialised desks share that corpus: a Greenbox Desk for guideline and policy questions, and a Market Desk for market queries. Per-document scoping keeps an answer inside the sources it should be reading.
- 03
Regenerate, copy, and feedback controls, plus a standing-by status indicator, so a broker can see the agent is grounded rather than guessing, and flag a bad answer in flight.
- 04
A custom agentic pipeline for loan officers. They upload a borrower's unsorted bank statements in one go; the pipeline reads them, merges them across banks and date ranges, classifies credits versus debits, and returns accurate totals, insights and loan product recommendations, even from sparse, disorganised data.
- 05
It runs in the background, so nobody waits on a screen. The loan officer gets an email when the analysis is ready, with filters by bank, account, statement range and description.
- 06
Extractions are named and linked to a client and loan number, so the output feeds downstream underwriting instead of dead-ending in a spreadsheet.
Recorded evidence
Only documented project facts appear here.
Bank statement review
- 1,000s
- Documents indexedInternal products + competitor knowledge
- 2
- Specialised desksGuideline policy · market queries
- 4 mo
- Engagement
The outcome
- Bank statement review went from two to three days of manual work to a 15–20 minute background process, with the loan officer emailed when it is done.
- Loan officers get totals, insights and a product recommendation from the same upload, instead of building them by hand.
More selected work
Next case studyHailo+Agentic agency infrastructure
Daisy+Unified agent platform
GiftologyGifting platform
MyAlex.aiAI voice CRM
Onyx EliteBrand transformation
AônPremium wellness brand identity
EternaGen LabsPeptide research e-commerce
Naveeda RealtyReal estate brand + digital presence
Project SetterVoice Agent + CRM Platform for the trades
Peak HumanLongevity platform: web + iOS, AI avatar, content at scale
Progressiv AIFintech: agentic AI operating layer across nine verticals
LucidSocialContent research, creation and distribution at scale