Projects

Featured project — localization.md

So that language stops being a barrier — a service that moves documents into another language with their original formatting and fixes what translation breaks.

Personal projects

Document platform PoC series — the predecessors of localization.md

Building a separate translator, editor, and checker for every format multiplies cost by the number of formats. I tested the hypothesis that handling every document through one common representation lets a single feature work across all formats — rebuilding it six times with a different architecture each time.

Hypothesis: convert PDF, PPTX, DOCX, and EPUB into a page-based common JSON (Unified Layout JSON); the web layer only reads, edits, and saves that JSON, and original formats are touched only at import and export.
Validation path: a UI-only PoC to check the user flow → a format adapter/exporter structure → an architecture splitting translation, conversion, viewer, and editor into separate services.
Design decisions it produced: separate format engines from the platform and make the common JSON the single source of truth; normalize reflowing documents such as DOCX and EPUB into page-fixed snapshots for review.
Extension: on the same representation, distinguish 'conversion' (keep the original) from 'generation' (use the original as source material — a PDF report into a slide deck, a manual into a training EPUB, tables into Excel).
Outcome: these decisions became the architecture of localization.md.

GitHub single-file operation-line challenge

Even a metric as simple-looking as 'lines of code' cannot be measured fairly without precise rules. I designed a challenge specification anyone can reproduce and verify under the same conditions.

Rules: fit as many distinct real operation lines derived from a = 1, b = 1 as possible into a single source file, with a final output of 2, in any language.
Fairness: every accepted line must connect directly or indirectly to the inputs; plain aliases, duplicates, blank lines, comments, and declaration-only lines are excluded, and a long line still counts as one.
Reproducible verification: runs within ten minutes on Azure Standard_D4s_v5 (4 vCPU, 16GiB, Ubuntu 24.04), prints 2 to stdout, exits 0. A line-based verifier and tests are included.
Handling rule changes: if GitHub's single-file size limit changes, classes are re-set and past records are kept with the limit, commit, and file hash of the time; records under different limits are compared by lines per MiB.
Contents: per-language examples (100+ lines each), identity-transformation patterns, a constraints JSON, and a language suitability table.

Data platform projects

PPAS backup & recovery / data lakehouse — Cut KT's PostgreSQL (PPAS) backup storage from 150TB to 7.53TB (about 95%) with Parquet on HDFS organized as Apache Iceberg tables.
Real-time analytics platform enhancement — Moved KT Skylife's real-time analytics platform for 7.5M households from CentOS 7 to a Rocky Linux 8 HA architecture with zero downtime and a 20-node scale-out.
MartDB generative AI data agent (GPT-4o) — A GPT-4o PoC for KT that answers natural-language questions over the PostgreSQL MartDB with RAG and text-to-SQL, cutting query turnaround from hours to seconds.
Airflow DAG batch scheduling — Moved KT's Crontab-based Spark and stored-procedure batches to Airflow DAGs on Docker with explicit task dependencies and web-based monitoring.
Insurance system migration from on-premises to AWS — Moved an insurer's core services from on-premises Docker hosts to AWS EC2 with security groups, least-privilege IAM, and EBS snapshot backups.
iOS/Android QA automation — Automated high-priority cases among Hana Bank's 3,300 manual app test scenarios with Appium and Pytest, running in parallel on three Android and three iOS devices.

Writing — on AI agents and architecture

All posts on LinkedIn

Harness engineering in practice: Twelve topics — repository structure, agent instruction files, linters, CI/CD, tests, permission and approval flows, execution loops, session handover — each with working examples grounded in Microsoft and GitHub documentation.
Spec-driven multi-agent orchestration: Starting from 'AI keeps improving, so why hasn't the way I work changed?', a summary of how organizations that actually adopted agents put them to work.
AIDE² — can an AI improve its own working environment?: An interpretation, from an Azure Data & AI practitioner's view, of a study in which agents ran 100 improvement loops over eight days without human intervention.
Why AI architects matter — through the structure of the IT industry: Where large companies outsource development and outside teams work implementation-first, design is the part most likely to be missing — and AI systems pay for it.
The AI pacing debate: A summary of industry leaders' positions around 'Pace the Frontier', viewed from a sovereign-AI perspective.
On redistributing leaked code: A critique of republishing leaked commercial source under a 'clean room' claim. The most-read post (about 27K impressions).
Everything is a Function: An essay on the tradition of modeling the world as functions — from Church's lambda calculus to programming languages and Java lambdas.
The missing multi-user collaboration in AI coding agents (report): Using public issues on Claude Code and Codex, an analysis of the gap that prevents several people from working in one agent session in real time.