Skip to main content

30-Minute Demo Script

Published 2026-05-14Updated 2026-06-306 min read

Goal: Prove within 30 minutes that the PoC is "not a concept but works in action" while conveying four values — 5-department perspectives, 3-BU integration, 4-type external signal fusion, governance.

0. Pre-Demo Preparation (5 min, outside demo time)​

CheckItem
OKTop-5 (S1·S2·S3·S5·S7) all respond within 1 min
OKAll 5 personas confirmed in sidebar order
OKcohort_tag badges (green/yellow/blue)
OKChatbot SSE trace panel functional
OKBackup scenarios (S4·S6·S8) validated once in advance
OKExternal data live (social · weather · economy · competitor)

1. Opening (3 min)​

"Today's demo is a PoC designed to solve the department data silo and perspective gap problem at LG H&H's Marketing Innovation Division. It is not a concept but a working demonstration — N=500~5,000 LG Members real data + 49.5K synthetic + 4 external signal types live all functioning."

Four key messages:

  1. Not a concept but a working demo — chatbot · graph · simulation all functional
  2. 5-department persona switcher — Brand Marketer / Insights / D&A / CRM · LG Members / MD
  3. 3-BU integration — Beauty + HDB + Refreshment unified into a single KG, visualizing member behavior crossing BUs · channels
  4. 4-type external signal fusion — social · weather · economy · competitor — signals invisible to first-party data alone

2. Top-5 Demo (20 min)​

Step 1 — S1. Natural-Language Semantic Search (4 min)​

"First, natural-language search. Let's enter as the Brand Marketer."

  • Persona: P1 (Brand Marketer)
  • Input: "Women in 30s, sensitive scalp, prefers organic, unpurchased in the last 90 days + Olive Young rating 4.5+"
  • Result: SKU card + Member card + first-party · Olive Young review combined card
  • Emphasis: "BM25 + embedding KNN + Reranker fused via RRF, showing not only first-party reviews but Olive Young · X external reviews as well. Simultaneous semantic + relational view via Neptune 1-hop graph."

Step 2 — S2. Persona Chatbot (5 min, demo highlight)​

"Now the chatbot. Starting with the MD persona."

  • Persona: P5 (MD · Channel Sales)
  • Input: "Top 5 channels where Perioe sales dropped this month and the cause?"
  • Right-side trace: "Autonomous invocations of channel_metrics_push → weather_join → social_trend_join"
  • After answer, switch persona to P1 (Brand Marketer)
  • Re-enter the same query
  • "Same data, but the answer tone, KPIs (channel→campaign), and next actions all change. This is the differentiator of department personas."

Step 3 — S3. Insight Cards (4 min)​

"Report automation + external signal fusion."

  • Persona: P3 (D&A)
  • Card: "Beauty BU monthly GMV + search trends + weather"
  • Click → AgentCore Code Interpreter generates matplotlib PNG + LLM comments
  • Switch persona to P5 (MD) → same card, different comments (GMV · inventory)
  • Emphasis: "Automated cards already combining the 4 external signal types — turn weekly reports into a one-click action."

Step 4 — S5. Campaign ROAS Simulation (4 min)​

"Marketing ROI simulation also works."

  • Persona: P1 (Brand Marketer)
  • Input: Budget 100M KRW / Su:m37 new product / VIP 50K
  • Result: Recommended channel mix donut + ROAS distribution violin
  • Change assumption (SNS +20%) → ROAS shifts immediately
  • Emphasis: "Separating campaign effect vs. trend effect by combining search trends · SNS responses — confidence-interval distribution rather than point estimates."

Step 5 — S7. Omnichannel Member Journey (3 min)​

"Finally, 3-BU integration."

  • Persona: P4 (CRM · LG Members)
  • Input: 1 high-value member ID / 90 days
  • Result: Owned mall (Beauty) → SNS ad (HDB) → Olive Young (Beauty) → CU (Refreshment) → VIP entry Swimlane timeline
  • Emphasis: "Beauty · HDB · Refreshment data once separated are now connected into a single KG — member behavior crossing BUs · channels on a single timeline."

3. Governance · Data Trust (3 min)​

3.1 cohort_tag Separation (30 sec)​

  • Demo result cards with green/yellow/blue badges
  • "Real / Synthetic / 4 external types specified on every card"

3.2 Guardrails (S8) (1 min)​

  • "Simulate SMS to 50K non-consenting members" → tool blocked + reason
  • "Member cust_001234 contact info" → PII masked
  • "Cosmetics campaign to an estimated-minor member" → automatically blocked

3.3 AWS Stack 1-pager (1 min 30 sec)​

  • Single architecture diagram
  • "Bedrock + Neptune + OpenSearch + AgentCore + Cohere — 8-week PoC feasible"

4. Closing (4 min)​

4.1 Summary (2 min)​

"What you've seen today is 5 of the 8 scenarios. S4 (persona matching + clustering), S6 (external signal fusion), and S8 (guardrails) operate on the same pattern."

4.2 Next Steps (2 min)​

StageDurationDeliverables
Discovery2 weeksFirst-party data inventory + accessibility review of 4 external types
PoC 8 weeks8 weeksTop-5 (real data N=500~5K)
Expansion+4 weeksFull 8-scenario set + guardrails · audit
Production transition+8 weeksLLMOps + ECS stabilization

"In 8 weeks, the same demo as today is feasible with LG H&H real data."

5. Frequently Asked Questions​

QuestionAnswer Highlight
Why KG instead of plain RAG?Plain RAG gives only semantics; KG RAG gives semantics + relations. S5 (attribution) and S7 (journey) are infeasible without KG
Korean morphology?OpenSearch Nori + Cohere embed-v4
Cost?$4K$6K/month for an 8-week PoC
What if first-party has Snowflake?Use Glue for Snowflake → S3 sync; same architecture applies
External data licensing?Public APIs primarily + PDPA · terms compliance review needed
Can PoC start with only some BUs?Yes. Beauty alone follows the same pattern
Difference at production transition?Same infrastructure + reinforced monitoring · LLMOps · guards

6. Demo Failure Fallbacks​

FailureResponse
Chatbot (S2) response delayCaptured screenshot → move to next scenario
Neptune query failureBypass via S1·S3·S5
External API (KMA · social) outagePre-cached fallback
Persona switcher bugURL direct (?persona=p1)

7. Post-Demo Deliverable Package​

  1. Demo recording (15-min edit)
  2. 8-scenario design specs (06-design-spec/)
  3. AWS single architecture diagram
  4. 8-week PoC milestones + cost estimate
  5. KG 25-class model (02-ontology.md)
  6. Data ingestion guide (04-data-sources.md)