Mobile Banking for Seniors, with AI-OCR
SonGeul
- Period
- Sep 2025 – Dec 2025
- Role
- PM · Frontend
Background
- Mobile finance usage among people in their 60s and older is 53.8%, more than 40 points below the 95%+ of people in their 20s to 40s (Bank of Korea, 2024). The senior population is expected to exceed 10M by 2030 (Statistics Korea 2024), and many seniors rely on family for financial tasks.
- Bank branch closures and ATM reductions have devastated rural financial access. The leading cause of erroneous transfers is 'account number input error.' Existing apps have high barriers for seniors due to complex authentication and security anxiety.
- Submitted to the 2025 KIITI Winter Academic Conference App Development Contest. I led a 6-person team (3 frontend · 3 backend) as PM across planning, AI architecture, and frontend development.
Execution
Product Planning & UX Design
- Designed an 'eliminate input' UX flow: capture (handwritten memo) → confirm (AI-OCR result) → approve (family 2-step auth) — simplified to 3 steps.
- Systematized supporting features under SAFE framework (Security · Assets · Family Connect · Education) and designed family-linked safety net structure.
AI System Design
- Designed a CLOVA OCR + Google Vision + fine-tuned model ensemble (weighted voting) to compensate for OCR accuracy limitations. Established a 3-stage anomaly detection roadmap (Rule → Statistical → ML: Isolation Forest, LSTM).
- Separated LLM roles into NLU (text semantic structuring) and XAI (risk explanation), and included a feedback loop for auto-learning from user correction data.
Business Model Design
- Sized the market directly from Statistics Korea senior population data and designed 4 revenue streams: B2C subscription / B2B licensing (leveraging the FSS senior service mandate) / financial brokerage / government grants.
Existing habit
Writing account numbers by hand — something seniors already do, used as input
① Capture
Weighted vote of CLOVA OCR, Google Vision, and our fine-tuned model
② Confirm
The user checks the recognized payee, account, and amount in large type
③ Approve
A family member confirms once more before the transfer runs
Sent
Typing the account number, the main cause of mistaken transfers, is gone
Starting from the premise that OCR can misread, the user and a family member each confirm once before the transfer, so authentication and security requirements stay intact.
Results
- Won Excellence Award at the 2025 KIITI Winter Academic Conference App Development Contest.
What I learned
- Instead of teaching seniors a new way, I turned their habit of jotting account numbers by hand into the interface. I judged that carrying a familiar behavior onto the screen asks far less of seniors than teaching them new controls.
- The design started from the premise that OCR has an error rate: an oversized confirmation step for the recognized number, plus ensemble correction. I judged that a confirmation step which stops a misread from becoming a wrong transfer is what builds trust in a financial service.