Skip to main content

S3. Category · BU Insight Cards

Published 2026-05-14Updated 2026-06-303 min read

1. URL Path

  • /insights

2. User Stories

P3 (D&A) — Builds the same report every week. Save time with automatic 4-type external signal fusion + LLM comments.

P5 (MD) — Sees a category sales card and wants to immediately ask "Why?" via the chatbot (S2).

3. Input UI

  • Card type selector (8 presets)
  • Period (week/month/quarter), BU/Category/Channel/Brand filters
  • "Generate Insights" button

4. Data Mix

DataSourceUse
First-party GMV (BU × Category × Month)Snowflake / Neptuneline/bar
Search trendsNaver · Google → SocialSignalDual axis
WeatherKMA → WeatherSignalDual axis
Competitor new productsCompetitorSignalannotation

5. Processing Pipeline

1. Card selection → openCypher / Snowflake aggregation
2. External signal join (date/region/category keys)
3. Result dataframe → AgentCore Code Interpreter
4. matplotlib + NanumGothic PNG
5. PNG + dataframe → Sonnet 4.6
6. Generate comments via persona system prompt
7. Card = PNG + comments + data table

6. 8 Card Presets

CardData
Monthly BU GMV trendBeauty/HDB/Refreshment comparison
Category × Channel GMVOwned mall · Mart · H&B · Convenience
Sunscreen GMV vs. Temperature · UVWeatherSignal combined
Beverage GMV vs. Outside temp · PrecipitationWeatherSignal combined
Search Trends vs. First-Party GMVSocialSignal combined
First-Party Impact After Competitor LaunchCompetitorSignal annotation
Membership Tier × GMVMembership × Order
Time-of-day Transaction DistributionOrderTransaction.order_at

7. Per-Department Comment Differences

PersonaComment example for the same card (Beauty monthly GMV)
P1 Brand Marketer"November Whoo +12% — estimated effect of pre-SMS campaign. Next campaign should..."
P2 Insights"Single-household persona's Beauty share rises to 25% — review new segment"
P3 D&A"November distribution +1.5σ vs. norm — temperature variable joint analysis required"
P4 CRM · LG Members"VIP point accrual rate 18% (+3pt) — tier retention campaign effect"
P5 MD · Channel Sales"Beauty turnover +0.4 — propose shelf expansion in department stores · Olive Young"

8. Guardrails

  • Mandatory source attribution for external data
  • No causal inference (correlation only) — LLM comment guard
  • Emphasize confidence intervals for small-sample categories

9. Demo Scenarios

  1. "Monthly BU GMV trend" card → demo comments across 5 personas
  2. "Sunscreen vs. Temperature" card → R² 0.62 scatter + order guide
  3. "Take this card to the chatbot" button → natural transition to S2