DongEun

01Workflow Automation2026.09

Overdue BTS Ticket Alert Automation

An Airflow pipeline that groups unresolved BTS tickets past their SLA by assignee and alerts them on Slack every morning

• ~10,000Unresolved tickets checked daily
• -21%Overdue count with business-day logic
• 21Unit tests, mutation-checked

• No read access to the ticket table behind the original dashboard SQL, so tickets come from JIRA REST (JQL) and are joined with SLA and store tables in Trino
• Counts elapsed time in business days instead of calendar days and skips weekends and holidays to avoid duplicate alerts
• Mentions assignees without any user lookup API by relying on Slack usernames matching JIRA email local parts
• Assignee workloads were highly skewed (median 4, max 597 tickets) and overflowed Slack's text limit; solved with a parent message plus threaded replies
• Separated core, task and DAG layers; verified with 21 unit tests that call the deployed code directly, plus mutation checks
• Traced the first production failure to a Slack bot swap made while the PR was open, and added a check for sibling-DAG config changes to the deploy checklist

• Python
• Airflow
• JIRA REST API
• Trino
• Slack API
• pytest

See project page

02Workflow Automation2026.03 ~ 2026.05

Abuser Appeal Email Automation

Automated appeal-email handling end to end — customer lookup, classification, template matching and reply

• 70 / weekRepetitive cases automated
• 7h / weekWorking time saved

• Connected Zendesk email, EOC customer status and abuse history to automate everything from lookup to reply
• Designed 11 inquiry types and a priority-based customer status classifier, with templates matched by reply count
• Human-in-the-loop flow — agents approve, reject or edit in Slack before the reply is sent
• SQLite history, duplicate-send prevention and unit/E2E tests for stable operation

• Python
• Zendesk API
• Slack API
• SQLite
• pytest

See project page

03Data Pipeline2026.06

CSAM Spend Aggregation & Reply Drafts

Aggregates spend per requested period and drafts a reply for agents to review

• Designed and built a pipeline connecting Zendesk – SharePoint – Trino/Hive – Google Sheets
• Split the period in each ticket into monthly and yearly ranges, aggregated them and generated reply text
• SQLite cache and pre-fetching cut repeated query cost and kept re-runs consistent
• Results are posted as Zendesk internal notes for final agent review to manage accuracy risk
• Compared manual replies with automated totals to find mismatches and data-source limits

• Python
• Zendesk API
• Trino/Hive
• Google Sheets
• SQLite

See project page

04Workflow Automation2026.04 ~ 2026.05

Weather Decision System & BW Caster

Bad-weather calls and notice images that took 7 people 4 hours a day — now one click

• 4h → 1minDaily manual work
• 6 · 8Templates · triggers

• Built a regional weather collection and decision system on public data and external APIs
• Linked KMA forecasts/warnings with Windy · ApiFlash to automate region selection → template choice → image generation
• Handled irregular API refresh cycles with Google Drive caching, 8 triggers, freshness checks and auto backfill/retry
• Kept tuning API quota, nationwide thresholds and representative-region logic on real data; integrated into the team AI Dashboard

• Python
• KMA API
• Windy API
• ApiFlash
• Google Drive API

See project page

05AI Evaluation2026.03

BW Toggle LLM Decision Validation

A quantitative experiment on whether an LLM can replace fixed-threshold bad-weather decisions

• 90.4%ON-bias root cause found
• 4 metricsAccuracy · Precision · Recall · F1

• Designed an experiment comparing threshold-based and LLM-based decisions
• Built an LLM pipeline with Pydantic structured output, batching, retries, checkpoints and token logging
• Evaluated LLM decision quality with Accuracy · Precision · Recall · F1 and mismatch cases
• Found why hourly aggregation skewed ON decisions to 90.4% and held accuracy near 53%
• Showed how aggregation unit and input representation affect quality, and set next experiments

• Python
• LLM
• Pydantic

See project page

06Monitoring2026.01 ~ 2026.02

CP-TO Watchdog Leading Indicator

A leading indicator and alert pipeline that catches operational issues before the lagging backlog does

• 13 minEarlier than existing trigger
• 100 → 1N+1 queries
• 3.5%False alarm rate

• Designed the Target On Duty formula to detect anomalies earlier than backlog
• Backtested on Gangnam data — 13 minutes earlier on average, AD10 diff +28.98%
• Pipeline for Grafana metrics → Google Sheets thresholds → comparison → 20-min cooldown → logging
• Cut per-region N+1 queries from 100 to 1 with SQLite composite indexes and batch reads
• Measured a 3.5% false alarm rate and defined tuning work toward under 1%

• Python
• Grafana
• Google Sheets
• SQLite

See project page

07Dashboard2026.02

CX Real-time Uploader & Ops Dashboard

Re-architected a monolithic fetch · render · send script into Airflow DAGs

• 10 minAuto-publish interval
• dozens → 4Core charts

• Split a single Selenium script's data fetching, image rendering and Slack sending into modules and Airflow DAGs
• Rendered 4 order charts, city panels and delivery-partner toggle alerts in parallel, published every 10 minutes
• Defined artifact lifecycle and re-run scenarios with Google Drive temp/queue folders and failure cleanup
• Cleaned up BW history schema and region data; designed Superset virtual datasets, filters and charts
• Reduced dozens of metrics to 4 core charts so the business could decide faster

• Python
• Airflow
• Selenium
• Slack API
• Superset
• Google Drive API

See project page

08Rule Engine2026.06 ~ 2026.07

OnePiece Mission Validation & SOP Rule Engine

Formalized natural-language SOPs into thresholds and AND/OR conditions in one Python rule engine

• Reorganized Mission · Segment SQL by purpose; fixed a missing leakage filter and ambiguous JOINs in the dashboard
• Verified metric reliability with a parity protocol comparing actuals to dashboard values, plus a KPI catalog
• Formalized andon, BW/GW toggle, booster and dispatch criteria as thresholds with AND/OR conditions
• Implemented a Pydantic input schema and a single rule engine; validated as a prototype, then integrated into OnePiece with test and demo UI
• Documented SOP terms and decision flows with a glossary, tables and Mermaid for validation and onboarding

• Python
• SQL
• Pydantic
• Mermaid

See project page

09Knowledge Base2026.07 ~ 2026.08

OnePiece Knowledge Base & Answer Viewer

Policies scattered across wikis, structured into a knowledge base for both people and MCP

• 11BPM rules consolidated

• Consolidated delivery-partner and 3PL pricing policies and 11 BPM rules from several wikis into a trigger–action structure
• Structured terms, formulas, issues and related projects so new members and MCP can reuse them
• Built an Answer Viewer — answers saved as local Markdown, viewed as HTML, feedback filed as GitHub issues
• Runs on premises, lowering privacy and security risk while closing the feedback loop

• MCP
• Markdown
• HTML
• GitHub Issues

See project page

10Service2026.08 ~ 2026.09

Disaster Map Service

A map that shows disaster alerts, CCTV, news and delivery status on one screen

• Prototyped disaster-alert collection, geocoding and Folium map visualization in Python
• Integrated the operational information into one always-on web view, updated with incremental collection and duplicate-run protection
• Designed one view combining CCTV, related news, delivery status and Slack alerts
• Improved news precision with a ±3-day filter and treated "no related articles" as a normal case
• Verified external API constraints locally and in GAS; researched alternative nationwide CCTV sources

• Python
• Folium
• Google Apps Script
• Slack API

See project page