Chongsun Yu | Portfolio

💬

cs-ai-chatbot

ArchivedAI PythonFastAPI

AI-powered CS chatbot prototype — evolved into cs-ai-dashboard (above). Claude API + RAG, 10 markets, fraud detection, smart escalation. Prototyping role completed.

cs-ai-chatbot/
├── app/
│ ├── main.py
│ ├── routers/
│ │ ├── chat.py
│ │ ├── escalation.py
│ │ └── metrics.py
│ ├── services/
│ │ ├── language_detector.py
│ │ ├── intent_classifier.py
│ │ ├── fraud_detector.py
│ │ └── response_generator.py
│ ├── rag/
│ │ ├── knowledge_base.py
│ │ └── template_store.py
│ └── models/
│ └── schemas.py
├── tests/ ← 161 tests
├── Dockerfile
└── .github/workflows/
└── ci-cd.yml

Key Features

🎯

Intent Classification

7 categories × 32 sub-intents, 90/10 hybrid

🛡️

Fraud Detection

9 Red Flag scoring (0~100) + Minna/Ethoca block

⚡

3-tier Smart Routing

L1 Escalate / L2 Draft / L3 Self-service (38%)

🌍

Multilingual Templates

30 templates × 10 languages, quality scoring

PythonFastAPIClaude APIChromaDBDockerPydantic v2GitHub Actions

🔴 PROBLEM

1,600+ monthly tickets across 10 markets in 10 languages handled manually

Inconsistent response quality — no fraud detection at intake

🟠 STRATEGY

Designed hybrid AI chatbot — 90% template consistency + 10% Claude AI flexibility

3-tier routing with RAG knowledge base and real-time fraud scoring

Full CI/CD: 161 tests → Docker build → staging deploy

🟢 RESULT

38% self-service target (~630 tickets/month automated)

9 Red Flag fraud detection operational with quality scoring (0.0~1.0)

Message Processing Pipeline

🖥️

cs-ai-dashboard

AIProduction PythonFastAPILive

Production AI-powered CS operations dashboard. FastAPI + Vanilla JS SPA integrated with Zendesk API. AI draft generation with Claude SDK, 9-market multilingual RAG, 12-pattern spam detection, 6-dimension quality scoring. Phase 1 Draft+Review operational.

toomics-ai-agent/dashboard/
├── main.py ← FastAPI app
├── routers/
│ ├── tickets.py
│ ├── actions.py
│ ├── drafts.py
│ ├── auth.py
│ └── auto_response.py
├── services/
│ ├── zendesk_client.py
│ ├── draft_generator.py
│ ├── claude_service.py
│ ├── rag_service.py
│ ├── spam_detector.py
│ ├── category_classifier.py
│ ├── language_detector.py
│ ├── quality_scorer.py
│ ├── killswitch.py
│ └── auto_responder.py
├── knowledge/ ← 9 lang RAG
├── static/ ← SPA (JS+CSS)
└── tests/ ← 1422 tests

Key Features

🤖

AI Draft Generation

95/5 rule — 95% source-based + 5% format only. Claude SDK/Proxy dual mode

🌍

9-Market Multilingual RAG

EN + FR/DE/ES/JP/PT/TH/CN/IT keyword-based section routing, EN fallback

🛡️

Spam Detection P1~P12

Phishing, legal impersonation, social media scam, security/sales spam

📊

4-Stage Category Pipeline

Intent → Content Scoring → Consensus → Post-correction (B2B detect)

🤖

🤖 L3 Auto-Response

Phase 1: 60s polling → 8-gate filter → AI Draft → Approve/Edit/Reject

PythonFastAPIClaude APIZendesk APIRAGJWTVanilla JSSPA

🔴 PROBLEM

1,600+ monthly tickets manually handled across 10 markets, 10 languages

No AI assistance — agents write every response from scratch

🟠 STRATEGY

Built production AI dashboard — agents review AI drafts, not write from zero

Invisible AI: customers never know AI exists. Zendesk UX unchanged

Phase 1 (Draft+Review ✅) → Phase 2 (Auto Send) → Phase 3 (Advanced)

🟢 RESULT

Phase 1 live in production — 1422 tests, Phase 1 Draft+Review operational

6-dimension quality scoring (Template/Safety/Action/Relevance/Complete/Length)

Live Dashboard Screenshots

Dashboard — Ticket List + KPI Cards + Charts

🤖 AI Draft Ready — Approve / Edit / Reject

L3 Settings — Master Control + Category & Market

🌐

ai-translate

AIProduction PythonFastAPILive

Air-gapped AI translation system for internal use. No external API calls — runs entirely on local SQLite vocab.db. Full-text search (FTS5) across 25,070 terms in 10 languages, 12 source integrations, 17-tier priority system. Changelog tracking for every edit.

ai-translate/
├── app.py ← FastAPI server (1,477 lines)
├── vocab.db ← 25,070 terms · 13.0MB
├── changelog.db ← full edit history
├── static/
│ ├── css/
│ │ └── style.css ← 473 lines
│ └── js/
│ └── app.js ← 855 lines SPA
└── templates/
└── index.html ← SPA shell

Key Features

🔍

FTS5 Full-Text Search

SQLite FTS5 index — partial matching, multilingual morpheme search across 10 languages

📦

12 Source Integration

Translation compilations, product glossaries, Confluence, app language packs — 17-tier priority

📋

Batch Translation

1~20 texts per batch, vocab.db pattern matching, auto language detection

📜

Changelog Tracking

Every add/edit/delete recorded in changelog.db — full audit trail

PythonFastAPISQLiteFTS5Vanilla JSSPAREST API

🔴 PROBLEM

10 language translations scattered across 12+ Excel sheets, Confluence pages, and PDFs

No centralized search — operators manually grep through files to find official terms

🟠 STRATEGY

Built air-gapped translation DB — no external API, no data leaves the server

Consolidated 12 sources with priority-based deduplication (37,839 → 25,070 terms)

SQLite FTS5 for instant search — no external search engine needed

🟢 RESULT

25,070 terms across 10 languages, 10 categories — single source of truth

CS Manual docs + FAQ 101 articles searchable in-app — no login required

Live Screenshots

Term Lookup — FTS5 search with autocomplete

Batch Translation — Auto detect → Target language

Docs — CS Manual + FAQ 101 articles, 10 languages

📈

global-cs-analytics

ANALYTICSMIT Jupyter

CS data analytics & interactive dashboards for multi-market operations. Ticket volume trends, SLA benchmarking, category analysis, and revenue save tracking.

global-cs-analytics/
├── notebooks/
│ ├── 01_ticket_volume_analysis.ipynb
│ ├── 02_response_time_benchmark.ipynb
│ ├── 03_category_deep_dive.ipynb
│ └── 04_revenue_save_tracking.ipynb
├── dashboards/
│ └── cs_dashboard.html
├── data/
│ └── synthetic_tickets_10k.csv
└── scripts/
├── data_generator.py
└── kpi_calculator.py

Key Features

🌍

Multi-market Volume Trends

10 markets, 11 languages, weekly WoW comparison

⏱️

SLA Benchmarking

FRT, resolution time, agent performance

🔬

9-stage Noise Filtering

82% noise removal → clean data pipeline

💰

Interactive Dashboard

Chart.js + Plotly KPI visualization

PythonJupyterpandasPlotlyChart.jsZendesk API

🔴 PROBLEM

Raw Zendesk data was 82% noise — no structured analysis possible

13K+ tickets across 10 markets with no cross-market pattern visibility

🟠 STRATEGY

Built 9-stage filtering pipeline with 4-tier category classification engine

Created interactive Plotly/Chart.js dashboards for KPI visualization

🟢 RESULT

82% noise eliminated — clean data pipeline established

13,400+ tickets across 7 months analyzed with proactive insights

📊 Data Pipeline

Raw Input

12,920

Noise (-82%)

-9,693

Clean Data

3,227

Total Analyzed

13,400+

🔧

bizops-automation-toolkit

EXECUTION MIT Python

Business process automation tools for global operations teams. Automates Zendesk reporting, chargeback tracking, Slack team digests, and Confluence documentation.

bizops-automation-toolkit/
├── zendesk-reporting/
│ ├── weekly_report_generator.py
│ ├── cs_kpi_analyzer.py
│ ├── template_manager.py
│ └── sample_data/
├── chargeback-tracker/
│ ├── sheets_sync.py
│ ├── alert_generator.py
│ └── sample_data/
├── slack-team-digest/
│ ├── collect_updates.py
│ ├── digest_generator.py
│ └── slack_webhook_sender.py
└── confluence-docs-sync/
├── manual_updater.py
└── template_generator.py

Key Features

📊

Weekly HTML Report Auto-gen

Zendesk API → Chart.js dashboard + KPI cards

🔄

Chargeback Real-time Sync

Google Sheets API → 30-column CSV/JSON

📱

Slack Team Digest

Weekly updates collection + webhook delivery

📝

Confluence Auto-updater

CS manual + category template generator

PythonZendesk APIGoogle Sheets APISlack WebhookConfluence APIpandas

🔴 PROBLEM

Manual reporting consumed 2 days/week — repetitive data collection across 4 platforms

Chargeback data siloed in Google Sheets with no automated tracking

🟠 STRATEGY

Built modular Python toolkit connecting 4 platform APIs independently

Each module runs standalone with auto-commit and sample data for portability

🟢 RESULT

Report generation 2 days → 1 hour (95% reduction)

Chargeback data real-time synced, team digest & docs fully automated

Automation Pipeline

🤖

CS_claw

AIProduction PythonTelegram Bot

Telegram-based AI operations assistant for server monitoring, Zendesk ticket lookup, audit execution, and deployment — all from a single chat interface. Claude Opus conversational AI with vision, SQLite session persistence, real-time health monitoring, and 15-pattern natural language command routing.

CS_claw/
├── bot.py ← single-file bot (1,342 lines)
├── config.py ← env + multi-project routing
├── session_store.py ← SQLite session persistence
├── doc_indexer.py ← document RAG indexer
├── episode_store.py ← conversation episodes
├── health.py ← PID lock + health check
├── data/ ← SQLite DBs
└── tests/

Key Features

💬

Claude Opus AI + Vision

Conversational AI with image analysis, streaming responses, auto-summarization at 20-turn threshold

🎯

Natural Language Routing

15 patterns — "차지백 현황", "티켓 131234" auto-routed without slash commands

🖥️

Server Operations

/health, /services, /logs, /deploy, /audit run — full ops from chat

🛡️

Resilience & Safety

fcntl.flock dedup, 409 backoff, 600s watchdog, rate limiting 10 req/min, confirmation for destructive ops

PythonTelegram APIClaude APISQLiteZendesk APIsystemdhttpx

🔴 PROBLEM

Server monitoring, ticket checks, audit runs require SSH + multiple CLI tools

No mobile-friendly interface — ops tasks blocked when away from desktop

🟠 STRATEGY

Unified Telegram bot — AI chat + server ops + Zendesk + reports in one interface

Natural language routing — no need to memorize commands

5-min health polling with auto-alerting on service down

🟢 RESULT

All ops from phone — ticket lookup, audit execution, deployment, log review

6 slash commands + 15 natural language patterns, Claude Opus AI with vision analysis

Processing Pipeline

🗂️

bizops-artifacts-portal

OUTPUTLive58 docs

Public documentation portal hosting 58 operational artifacts across 9 categories. CS manuals (EN/KO), FAQ (10 languages), ToS (10 languages), chargeback templates, AI dashboard docs, analysis reports, and weekly archives.

bizops-artifacts-portal/
├── index.html
├── cs-operations/ ← 10 docs
│ ├── cs_manual_EN.html
│ ├── cs_manual_KO.html
│ ├── fraud_analysis.html
│ ├── cs_infographic.html
│ ├── cs_flow_diagram.html
│ └── weekly-reports/ ← W1~W4
├── chargeback/ ← 3 docs
│ ├── operation_manual.html
│ ├── template.xlsx
│ └── flow_diagram.html
├── ai-dashboard/ ← 6 docs
│ ├── policy.html
│ ├── quickstart.html
│ └── l3_engine.html
├── ip-protection/ ← 1 doc
│ └── ip_manual.html
└── dashboards/ ← 2 docs
├── ticket_dashboard.html
└── dashboard_manual.html

Categories

📞

CS Operations (10 docs)

EN/KO manuals, fraud report, flow diagrams, weekly reports

🖥️

AI Dashboard (6 docs)

Policy, quickstart, L3 engine, roadmap, settings

💳

Chargeback (3 docs)

8 PG dispute manual + evidence form V5 + template v2

📊

Dashboards & IP (3 docs)

Ticket dashboard, manual, IP protection guide

GitHub PagesHTML/CSSGitHub ActionsClaude AI@media printConfluence

🔴 PROBLEM

Operational docs scattered across Confluence, Sheets, and local files

No single source of truth for 10 global markets' operations

🟠 STRATEGY

Built centralized GitHub Pages portal with CI/CD auto-deployment

5 categories, EN/KO bilingual, print-optimized for PDF export

🟢 RESULT

58 documents publicly accessible — all authored solo

Single source of truth for 10 global markets with auto-deploy on push

📁 Documents by Category

AI Dashboard

6

CS Ops

10

Chargeback

3

Dashboard

2

IP Protection

1

Portal Screenshot

BizOPS Artifacts Portal — 58 docs, 9 categories

⚡

nexus-dashboard

DevOpsProduction TypeScriptNext.jsLive

Claude Code workspace management dashboard. Next.js 16 + shadcn/ui — manages agents, skills, hooks, and project instructions across multiple repositories. JWT auth, CRUD operations, audit logging, real-time monitoring, and export/import for configuration portability.

nexus-dashboard/
├── app/ ← Next.js 16 App Router
│ ├── overview/ ← KPI cards + project list
│ ├── agents/ ← agent CRUD + routing
│ ├── skills/ ← skill registry + hot reload
│ ├── hooks/ ← hook toggle + event config
│ ├── instructions/ ← CLAUDE.md editor
│ ├── projects/ ← multi-repo management
│ ├── monitor/ ← real-time monitoring
│ └── settings/ ← config + export/import
├── components/ ← shadcn/ui components
├── lib/ ← auth + API utilities
└── pm2.config.js ← PM2 deployment

Key Features

🧩

8-Page Management Console

Overview, Agents, Skills, Hooks, Instructions, Projects, Monitor, Settings

🔐

JWT Auth + Audit Log

Secure authentication with full audit trail — every create/update/delete logged

🔍

Search + Highlighting

Global search across agents, skills, hooks — keyword highlighting in results

📦

Export / Import

Full configuration portability — export as JSON, import to new workspace instantly

TypeScriptNext.js 16shadcn/uiTailwind CSSJWTPM2REST API

🔴 PROBLEM

Claude Code configuration scattered across CLAUDE.md, agents.yaml, hooks — no unified view

Managing 7+ projects × dozens of agents/skills requires manual file editing

🟠 STRATEGY

Built visual management dashboard — CRUD for all Claude Code primitives

Project-scoped filtering — switch context without navigating file trees

shadcn/ui component library — consistent, accessible, responsive UI

🟢 RESULT

8-page dashboard managing 7 projects, 40+ agents, 60+ skills from a single UI

Full audit log + export/import for configuration portability across environments

Live Dashboard Screenshots

Overview — KPI cards, project list, quick actions

📖

payment-ops-playbook

EXPERTISECC BY 4.0Docs

Open knowledge base for payment gateway operations. Covers PG fundamentals, cross-border settlements, fraud & chargeback management, and market localization from hands-on experience.

payment-ops-playbook/
├── 01-pg-fundamentals/
│ ├── payment-flow-overview.md
│ ├── 3-party-vs-4-party-system.md
│ └── card-brand-acquiring.md
├── 02-cross-border-payments/
│ ├── cross-border-settlement.md
│ ├── fx-risk-management.md
│ └── aml-kyc-compliance.md
├── 03-fraud-and-chargeback/
│ ├── fds-rule-design.md
│ ├── chargeback-dispute-process.md
│ └── fraud-pattern-analysis.md
├── 04-localization/
│ ├── thailand-pg-setup.md
│ └── multi-market-operations.md
└── diagrams/
├── payment-flow.mermaid
└── settlement-cycle.mermaid

Chapters

💳

PG Fundamentals

Payment flow, 3-party vs 4-party, card brand acquiring

🌏

Cross-border Payments

Settlement, FX risk, AML/KYC compliance

🛡️

Fraud & Chargeback

FDS rule design, dispute process, pattern analysis

🇹🇭

Market Localization

Thailand PG setup, multi-market playbook

Visa / MastercardNHN KCPCross-borderFDSAML/KYCMermaid

🔴 PROBLEM

PG operations knowledge siloed in individuals — no structured handoff

No onboarding guide for cross-border settlement or fraud operations

🟠 STRATEGY

Documented 10+ years of PG experience into 4 structured chapters

Added Mermaid diagrams for visual payment & settlement flows

🟢 RESULT

Comprehensive playbook covering 4 core domains with visual flows

Reusable reference for PG partnerships, FDS design, multi-market expansion

🎭

avatara-nft-platform

BLOCKCHAINNX3 Games Web3

Blockchain-based digital avatar NFT platform designed at NX3 Games (NEXON). Led end-to-end service design — NFT minting pipeline, marketplace trading, tokenomics, and GameFi integration on Polygon network.

avatara-nft-platform/
├── contracts/
│ ├── AvataraNFT.sol
│ ├── Marketplace.sol
│ └── TokenRewards.sol
├── avatar-builder/
│ ├── trait_generator.py
│ ├── metadata_engine.py
│ └── rarity_calculator.py
├── marketplace/
│ ├── listing_service.ts
│ ├── auction_engine.ts
│ └── royalty_splitter.ts
├── gamefi-integration/
│ ├── staking_module.sol
│ └── play_to_earn.sol
└── docs/
├── tokenomics.md
└── whitepaper.md

Key Features

🎨

Generative Avatar Minting

Trait-based generative system with rarity tiers

🏪

NFT Marketplace

Listing, auction, royalty split on Polygon

🎮

GameFi Integration

Staking rewards + Play-to-Earn mechanics

📊

Tokenomics Design

Supply control, burn mechanics, reward distribution

SolidityPolygonNFTWeb3.jsIPFSReactTypeScriptHardhat

🔴 PROBLEM

Game assets locked in centralized servers — no true player ownership or cross-game portability

No secondary market for in-game items — zero residual value for players

🟠 STRATEGY

Designed NFT-based avatar system on Polygon — low gas, high throughput for game assets

Built marketplace with royalty splits (creator + platform) and P2E staking mechanics

Managed tokenomics — supply, burn, reward cycles aligned with game economy

🟢 RESULT

End-to-end NFT platform designed — minting → marketplace → GameFi full lifecycle

Web3 partnership ecosystem established with multiple blockchain partners

NFT Lifecycle Architecture

💱

global-mass-remittance

NEW BIZNHN KCP Fintech

Designed and launched a cross-border multi-currency mass remittance platform for global partners. Developed KRW-KRW and USD-USD localized bulk transfer services for Payoneer, Uber, and Airbnb — enabling mass bank transfers to sub-merchants without a Korean PG license. Generated ₩200M+ monthly fixed revenue.

Monthly Fixed Revenue

₩200M+

Minimum guaranteed / month

Revenue Model

Fixed + %

Fixed KRW per transfer + commission rate

Transfer Frequency

Daily

Daily & weekly cross-border USD settlement

global-mass-remittance/
├── 01-cross-border-usd/
│ ├── usd-virtual-account-design.md
│ ├── bulk-transfer-api-spec.md
│ ├── settlement-cycle-daily.md
│ └── revenue-model-analysis.md
├── 02-localized-remittance/
│ ├── krw-krw-domestic-transfer.md
│ ├── usd-usd-fx-settlement.md
│ └── sub-merchant-onboarding.md
├── 03-partner-integration/
│ ├── payoneer-integration.md
│ ├── uber-settlement-flow.md
│ ├── airbnb-host-payout.md
│ └── partner-compliance-kyc.md
├── 04-van-banking/
│ ├── van-partner-contract.md
│ └── banking-api-integration.md
└── diagrams/
├── remittance-flow.mermaid
└── settlement-cycle.mermaid

Key Deliverables

💵

Cross-border USD Mass Remittance

USD virtual account-based bulk transfer with daily/weekly settlement

🔄

KRW-KRW / USD-USD Services

Bulk bank transfers for partners without Korean PG license

🤝

Global Partner Acquisition

Payoneer · Uber · Airbnb onboarded for Korean market settlement

🏦

VAN Banking Infrastructure

Mass bank transfer execution and reconciliation

🔒

Partner Lock-in & Revenue Defense

Remittance as value-add → partner retention + new acquisition

Cross-borderUSD Virtual AccountMass RemittancePayoneerUberAirbnbVANKRW-KRWUSD-USD

🔴 PROBLEM

Global e-wallet providers (Payoneer, Uber, Airbnb) needed Korean sub-merchant settlement but couldn't obtain PG licenses

NHN KCP revenue relied solely on PG processing fees — no defense against partner churn

No infrastructure for cross-border multi-currency bulk transfers to Korean recipients

🟠 STRATEGY

Designed USD virtual account-based cross-border bulk remittance platform with daily/weekly settlement

Developed KRW-KRW & USD-USD localized mass transfer services for partners without Korean PG licenses

Partnered with domestic VAN providers for banking API integration and automated reconciliation

Positioned remittance as value-add for existing PG partners → lock-in + new acquisition channel

🟢 RESULT

New platform launched — ₩200M+ monthly fixed revenue with per-transaction commission

Onboarded Payoneer, Uber, Airbnb as remittance clients for Korean market settlement

Dual revenue model: fixed KRW fee + % commission on each cross-border remittance

Existing partners locked in; remittance became a new partner acquisition channel

Mass Remittance Platform Architecture

📊 Service Impact Metrics

Monthly Rev.

₩200M+

Global Partners

3+

Transfer Types

KRW+USD

Settlement

Daily

Chongsun Yu
PM who builds × Product Owner

Why I Built Everything Below

The Problem

No One's Coming

Data Said Yes

Built It on My Own Dime

The Pivot

Production Quality

Knowing Where to Stop

Built to Solve Real Problems

12+ Years in Global Operations

Team Lead, BizOps

Part Lead, Blockchain Business

Part Lead, Global Payment Division

IT Director / Team Lead

Account Manager, Global Payment

Australian National University

Tools & Domain Expertise

Domain Expertise

Tech & Tools

Let's Connect

Chongsun Yu PM who builds × Product Owner

Why I Built Everything Below

The Problem

No One's Coming

Data Said Yes

Built It on My Own Dime

The Pivot

Production Quality

Knowing Where to Stop

Built to Solve Real Problems

12+ Years in Global Operations

Team Lead, BizOps

Part Lead, Blockchain Business

Part Lead, Global Payment Division

IT Director / Team Lead

Account Manager, Global Payment

Australian National University

Tools & Domain Expertise

Domain Expertise

Tech & Tools

Let's Connect

Chongsun Yu
PM who builds × Product Owner