THU · AUG 20 · DAY 2

What’s next

13:18 KST

NEXT MOVE 42 min

SA

SHINHAN CARD · 신한카드

Sunghee Ahn

Head of AX Research Institute

14:00 TRACK A

From RDB to GraphDB: Unlocking Relationship Intelligence in Financial Data

Why now

31M customer profiles + graph-data work. Exact relationship-preservation wedge; pain still needs validation.

NEEDS ATTENTION

Don’t drop the thread

2

IF THE PLAN BREAKS

Two good alternatives

33 QUALIFIED · 12 P1 ROUTES

Targets

QUICK RECEIPT

Capture

00:00

↯

Fast first. Clean later.
Partial identity and rough notes are enough to save.

“I may be framing this wrong—how does your team handle this today?”

4 OPEN · 1 DUE SOON

Follow-ups

JP Jiyoon Park Hanwha Life · AI Platform DUE 3H

Captured Structured Review Send

ORIGINAL FIELD NOTE 14:07 · MING

Runs internal claims copilot. Real claims docs can’t leave prod so eval set is fake and misses Korean medical terms. Asked for technical brief + intro to security lead. I said today.

AI STRUCTURE 3 certain · 1 check

Problem: Production claims documents are unavailable for realistic evaluation.
Current workaround: Artificial evaluation set with weak Korean medical terminology.
Promise: Send technical brief today; request security-lead introduction.
Qualification CHECK: Likely evaluation owner; budget and implementation timeline unknown.

OWNERMing SENDERMing CHANNELEmail DEADLINE17:00 KST

REVIEWED DRAFT

Jiyoon — good talking after the session. You mentioned the claims copilot team is evaluating against a synthetic set because the real documents cannot leave production, and that the Korean medical language is where the test set falls short.

I’m sending the technical brief I promised. The relevant piece is keeping the same patient and claim entities consistent across the documents while replacing the sensitive values, so the evaluation still behaves like the production workflow.

Happy to walk your security lead through how it runs inside the customer environment. Worth including them on a short technical session next week?

Ming

HubSpot note + task ready after send

MH Minho Hwang · Kakao Identity incomplete · finish original note DRAFT

YC Yongmin Choi · Woori Wedbush warm thread · Ben owns WARM

PROTOTYPE STATE CHECKS

AI structuring pendingOriginal note already safe AI PENDING

HubSpot update failedRetrying; field note remains safe RETRYING

Ben + Ming edited offlineBoth versions preserved for review CONFLICT

Security-lead introductionOwner Ben · deadline missed by 2h OVERDUE

Warm introduction sentContinue in the introduced thread WARM SENT

Cold route eligibleNo person or account hold COLD OK

DAY 2 · LIVE SYNTHESIS

What changed

11
convos

STRONGEST SHIFT

The blocker is evaluation fidelity—not simply access to training data.

Four teams described test or evaluation environments that fail to reproduce the relationships and language in production data.

03 REPEATED

“Synthetic data” is heard as fully generated rows.

Lead with consistent entity replacement, then explain the broader category.

01 EARLY SIGNAL

Korean medical terminology may be a product-proof gap.

One conversation only. Validate before changing the roadmap or message.

CHANGE TOMORROW

01
Ask about evaluation and test environments before model training.
02
Say “replace the sensitive entities” before saying “synthetic data.”
03
Probe whether Korean-language fidelity breaks current workarounds.