Richard Dufaur

I direct an AI-assisted data system that turns the Korean music industry into structured, verifiable intelligence.

flow. is a weekly, data-driven read on the business of K-pop, and this is the system behind it. I own the editorial doctrine and the final call, the AI does the collection and the first pass.

Flagship project

flow.

A weekly English newsletter on the K-pop industry, written from a business and data angle, not fan content. It reads the Korean music market the way you'd read any other sector: revenue, filings, charts, exports.

Every issue runs on a production data pipeline with AI in the loop: primary Korean press and regulatory filings collected, resolved, and cross-checked before a single figure is published.

I run it end to end. I set the editorial doctrine, direct the AI-assisted collection and analysis, and decide what ships. The authority comes from access to the data and the primary Korean-language press, not borrowed encyclopedic fandom.

23
Weekly issues
~56
Database tables
~2,400
Indexed documents
Share prices of HYBE, JYP, SM and YG indexed to 100 over one year — a chart built from the flow. pipeline's KRX data
Built by the flow. pipeline — KRX closes, four labels indexed over one year.

Architecture

How the pipeline works

Raw sources become structured, sourced facts before they reach a reader. Five stages, fully automated.

01
Sources
KR press, filings, markets
→
02
Collection
Crons, healthchecks
→
03
Database
SQLite (WAL), ~56 tables
→
04
Analysis
Facts, time series, tests
→
05
Publication
flow. weekly

Data sources

Tier-1 Korean press and international business media, plus financial and market data down to regulatory filings and customs exports.

NaverDaumNewsen Music Business WorldwideBillboard VarietyKorea Herald DART filingsKRX Circle ChartKOSPI / KOSDAQ USD/KRWKorea Customs

Entity resolution

A de-duplicated reference of K-pop entities (groups, agencies, artists) with a multilingual alias table so names resolve across languages and spellings.

에스파  →  aespa

Structured fact store

Every published number is stored as a record (metric, value, unit, period, source, confidence) that builds cumulative time series rather than a pile of articles.

metricvalueunit periodsourceconfidence

AI-assisted tooling

A custom CLI drives curation and issue prep. An AI hypothesis tester takes a claim in plain language, gathers evidence across every data surface, and returns a verdict. Semantic search runs over the full document store.

kpulse CLIAI hypothesis testing semantic searchcollection crons

Verification doctrine

Nothing ships unsourced

The system is built around fact-checking gates. A written editorial doctrine and mechanical checks stand between the data and the reader, every week.

  • Human in the loop. The AI gathers, drafts, and flags. I own the doctrine and make the final publication call.
  • Primary source, traced. Every figure links back to a primary source before it can be published.
  • Verdicts, not vibes. A hypothesis returns supported, partial, contradicted, or insufficient data, never a guess.
  • Cumulative memory. Facts accumulate as time series, so each issue builds on a verifiable record.
  • Automated integrity. Collection crons, pipeline healthchecks, and email alerts keep the data current and catch failures.

Background

Foundations

Eight years in cultural mediation, then a data-analyst training. The training projects below cover the fundamentals: SQL, Python, Power BI, machine learning.

Early projects, data-analyst training 10 projects · click to expand

Coursework deliverables, kept for reference. The production system above is where the real work is now.

P2E-commerce sales analysisView deliverables
P3SQL relational databaseView deliverables
P4Public-health study (R)View deliverables
P6Retail inventory optimizationView deliverables
P7Power BI dashboardView deliverables
P8Gender-equality indicators (GDPR)View deliverables
P9Time-series correlationView deliverables
P10Spotify music-evolution dashboardView deliverables
P11Market study (PCA)View deliverables
P12Fraud detection (ML)View deliverables