Turning business problems into tools people actually use
Jaewoong Yoon (윤재웅). Senior Manager in the DX Innovation Office at POSCO International, company-wide instructor for Power BI, Power Apps and Power Automate, and Technical Director of the Korea Microsoft Power Platform Association. I have taught Python data analysis to about 1,600 employees, and I hold an MBA with a thesis on forecasting Korea's wholesale electricity price with LSTM.
1,600employees taught Python data analysis
21new-hire and DX program classes taught as lead instructor in 2026 · rated 4.8/5
15business DX and RPA builds supported (2025–2026)
Build it fast, then make it stick in the organization
I started in 2003 as an electrical and controls engineer for display and electronics equipment, joined POSCO DX in 2008, and have worked across ERP, data analysis, RPA and visualization in the POSCO Group ever since. At POSCO Energy (now part of POSCO International) I worked on a company-wide process diagnosis, then joined the PI task force it led to, owning the plant maintenance and quality modules in the move from Oracle ERP to SAP ERP. I then led SAP operations for five years, including audits, closing issues, a plant expansion rollout and S/4HANA planning, and separated the IT systems of the Korea Fuel Cell spin-off. More recently I led the PMO for an overseas subsidiary's ERP rollout and the ERP selection for another.
Today I plan and deliver company-wide DX, and I spread data analysis, automation and visualization skills through a course program I designed myself.
When someone in the business says "I wish we had a tool for this," I use generative AI and vibe coding to show a working prototype within minutes. I aim to be the person who builds the solution and gets it adopted, not just the one who explains the technology.
ProjectsSelected
Selected work
ERP rollout PMO for subsidiaries in Poland and India
2024–present · POSCO International
Role
PMO for the ERP rollout at the motor core subsidiary in Poland, a 23-person project team; ERP solution selection and PMO for the motor core subsidiary in India, where the rollout mirrored Poland and the PMO work centred on issue management.
Work
Designed the competitive bid, wrote the RFP, built the proposal scorecard with weights by criterion and by department, and, for India, compared candidates such as QAD and TCS iON (the cloud ERP Tata Consultancy Services offers Indian small and mid-sized manufacturers) against about 133 functional requirements in purchasing, production, costing and finance.
Result
Poland: contract with QAD in October 2024, finance and purchasing modules live in May 2025, followed by production in July 2025. India: contract signed with QAD in May 2026, rollout started the same month and completed in October 2026.
Phase 2
Scoping, as a second-phase project, additional management accounting development for the three motor core subsidiaries (India, Mexico, Poland) to fit head office’s global consolidated profit and loss management program.
ERP
PMO
RFP · bid evaluation
Overseas subsidiaries
Q-Procure Assistant: AI agents for strategic purchase timing
2025.11 · POSCO Group Hackathon finalist
Idea
A team of AI agents that supports the whole purchasing flow: item registration from a photo, stock depletion forecasting, target price calculation from exchange rates, supply chain risk and market prices on the web, an automatic purchasing strategy report, and an alert when the timing is right.
Role
Team designer: shaped the service flow and built agents.
Result
Reached the finals of the POSCO Group Hackathon (November 6–7, 2025).
AI agents
Procurement
Hackathon
Building DX skills at an overseas subsidiary: Global Mobility Program
2026.09–2026.11 · Work mentor
Context
A 13-week head office rotation for a local employee of the Indonesia trading subsidiary, the first of its kind, so the subsidiary can run its own DX work.
Work
Designed the program and taught Power BI, Power Apps and Power Automate in English, then coached two real projects. Daily accounts receivable reminders: SAP receivables pushed to Teams and Outlook, cutting weekly work per person from one hour to under 15 minutes with zero manual reminders, with users set to grow from 6 to 25. Single View dashboard: SAP sales, trade statistics and location data combined in Power BI across three views.
Now
Migrating the AI agents the subsidiary had already built onto the company’s vibe coding sandbox (Track 2) and rebuilding them in a managed environment; final presentation at the end of November.
Global talent
Taught in English
Power Platform
SAP
IT carve-out for a spin-off: SAP to Douzone iCUBE
2019.11–2020.08 · Korea Fuel Cell (secondment)
Work
Separated the IT systems of Korea Fuel Cell when it was spun off from POSCO Energy: moved the SAP-based business environment to seven Douzone iCUBE cloud modules (accounting, HR, time and attendance, sales, purchasing and materials, trade, service management), set up an independent IT operation including the EP groupware, supported data and user migration, and helped set up the office environment, including network equipment and printers.
People
Trained the new company's first IT team, one for operations and one for administration, so it could run its IT on its own.
Result
Business continued without interruption after the spin-off, and the new company had its own staff ready to run its IT.
Carve-out
SAP
Douzone iCUBE
7 modules
Data migration
SAP operations lead, plant expansion PM and next-generation ERP planning
2014.07–2019.10 · POSCO Energy
Work
Owned all SAP modules (SD, MM, FI, CO, PS, QM): month-end closing issues, change approval through ITSM with spec reviews, and monthly authorization reviews that passed SAP audits. Handled SAP license audits as a certified software asset manager. Led the SAP rollout for a new cell plant (2015, SD, FI, PM, MM), planned the move to SAP S/4HANA, and ran the standard process management tool (Real BPA).
Result
Division head's commendation (2018) for cost and process improvement; Employee of the Year (2019).
SAP operations
Audit
ITSM
S/4HANA
Company-wide process diagnosis that launched a PI task force
2012.04–2014.07 · POSCO Energy
Work
First round: investment, production, purchasing and R&D processes, centred on the new investment business. Second round: execution processes for power plant construction, operation and maintenance, to raise profitability. I gathered the data, VOC (voice of the customer) and VOB (voice of the business), and wrote the reports.
Result
The findings were presented to the POSCO chairman, a meeting I attended, and led to a new PI task force in October 2012, where I owned the plant maintenance (PM) and quality (QM) modules in the Oracle-to-SAP ERP transition.
Record
The full story of PI3.0, the company-wide process innovation program run from 2012 with every employee taking part, was published in 2015 as the book POSCO Energy, Meeting Tomorrow First, which I co-wrote (Energy Economy News, May 4, 2015). Proceeds went to charity.
Process diagnosis
VOC · VOB
PI
Oracle → SAP
PM · QM
PI3.0
The PI3.0 book, published 2015 · Author · far left · Photo: Energy Economy News, May 4, 2015
Automated denied-party screening for export customers
2026 · POSCO International
Problem
Employees each searched the government strategic trade control system by hand whenever they needed to check a customer, repeating the same checks.
Work
Designed and built an RPA bot that screens all of the roughly 20,000 customers registered in SAP and posts the results on the company notice board.
Result
Manual one-off searches disappeared; anyone can see the screening result for every customer in one place.
RPA
SAP
Web data collection
Citizen development support: 15 builds
2025–2026 · POSCO International
2026
8 business DX projects: 4 Power BI dashboards (steel results, electrical steel sales, import statistics by country, raw material trading) and 4 automations.
2025
7 self-built RPA flows coached to completion, such as exchange-rate crawling into SAP, oil price and news digests, and AI-drafted reports from IEA publications.
Result
Business users now build and run their own automations without filing development requests.
Power BI
Power Automate
Generative AI
DX Contest evaluation framework for European subsidiaries
2026
Work
Designed five criteria (impact, scalability, innovation, technical difficulty, feasibility) and a scoring scheme to evaluate 12 DX projects from 5 European subsidiaries objectively.
Personal apps built by vibe coding: Plan & Do planner, Coffee counter
2026 · personal · Claude
Planner
Started as a 61-spread 2027 plan-vs-actual PDF for Samsung Notes, generated in Python (ReportLab, korean_lunar_calendar) with linked yearly, monthly and weekly pages. Then rebuilt as an installable app where one entry appears in both the monthly calendar and the 30-minute weekly timetable, with plan and actual side by side.
Coffee counter
One tap logs a drink; it totals caffeine, sugar and creamer against a 400 mg daily limit and charts 7, 30 and 90-day trends.
How
Built without hand-writing code, by describing screens and rules to Claude and refining them.
Tech
Offline-first PWA (manifest, service worker, localStorage), Firebase Authentication and Cloud Firestore with offline cache and per-user security rules, hosted on GitHub Pages. Sync merges per item (each to-do, goal, time block and note), keeps the latest edit and stores the other side in a conflict log; offline conflicts were tested with two browsers. Open the app ↗
Weekly view · sample data
TeachingTalks · Books
Teaching, talks and writing
How I design course books: industry → company → department · I design course books and practice datasets by narrowing from the broad to the specific, so learners practice on screens that look like their own work and apply it the next day.
Industry
Reflect the deal structure of the industry, such as trading, manufacturing or energy. Example: 600 three-country trade transactions for consolidated profit and intercompany analysis.
Company
Match the company's business structure and systems. Example: results data modeled on steel, energy and food divisions; SAP OData integration.
Department
Practice with the metrics and screens the team sees every day. Example: results, equipment monitoring and subsidiary P&L dashboards.
When
What
2026.09
Speaker, Modern Work Conference Seoul 2026 (Modern Work Community, at Microsoft Korea). "Become a Power BI expert in no time with Copilot." Event page ↗
2025–2026
Instructor, DX Associate Expert Program at POSCO International: beginner and intermediate Power BI and Power Apps for employees, in an official in-house program whose completion counts toward promotion. 15 classes in 2025 (new-hire training and the DX program) as lead or assistant instructor; 21 classes in 2026 (new-hire training and the DX program) as lead instructor, with an average learner rating of 4.8 out of 5. All course books written by me (113 to 206 pages each).
Why I teach
After years of building and running systems, self-build support projects were the first time business users told me "this really helps." That is why I spent my own time writing the DX Associate Expert Program course books and keep coaching employees.
2020.09–2022.12
Seconded instructor, POSCO Human Resources Development Center: Python data analysis in the POSCO "New Collar" program, a popular course because completion counted toward promotion; a two-day class of about 30 people in each of the first three weeks of every month, with re-assessments in the fourth week, about 1,600 learners in total. Certified A-grade instructor (2022.03). When recording an online R course requested by an outside company, I automated the on-screen code typing and execution with Python RPA, which made producing the videos far easier.
2021 · 2022
Python and RPA classes for Pohang Jecheol Technical High School students, as part of the company's community education support.
In progress
Book: "Power Platform Automation with Copilot" (working title, about 400 pages): collect with Power Automate, analyze with Power BI, act with Power Apps, with generative AI at every step. Sample chapters (Korean, PDF) ↗
2015
Co-author:POSCO Energy, Meeting Tomorrow First, the book on the company-wide PI3.0 process innovation program. Sold in Korean bookstores; proceeds went to charity.
Materials
In-house courses currently delivered: Power BI and Power Apps, beginner and intermediate. The other materials are finished course books prepared for upcoming courses. Downloadable handouts in Korean: vibe coding with Python and Claude Code, SAP CDS View and OData to Power BI, Power BI dashboard labs, process mining with a Power Automate Task Mining lab, an enterprise Claude adoption and security checklist, and attack surface management for vibe coding. See the list ↗
NewsAI · Power Platform
Power Platform news · 2026.11.07
Dataverse moves into AI tools, Power BI adds visuals, Architecture Center standardizes access
Dataverse in more AI tools (Oct–Nov 2026)A Dataverse plugin for Copilot Cowork, an n8n node, and plugins for coding agents such as GitHub Copilot and Claude. Agents stay within each user's existing permissions.
Power BI October 2026 updateGantt chart and dropdown list slicer (preview), plot area background color, custom visuals through Skills for Fabric, continuous capacity usage reporting for semantic models, and a mobile page indicator.
Architecture Center, November 2026New reference architectures for secure multi-account access in Power Pages and for standardizing Dataverse access control with Microsoft Entra security groups; September added SAP order acknowledgment matching with Copilot Studio: Power Automate captures supplier confirmations from mail or EDI into Dataverse, an extraction agent normalizes PO number, lines, quantities, prices and dates, an orchestrator pulls the PO from SAP through the SAP connector, a matching agent applies exact-match or tolerance rules, matches are written back to SAP automatically, and discrepancies go to buyers in Teams, with every outcome logged in Dataverse and least-privilege Entra service accounts.
Takeaway · As agents read and write business data directly, access design comes first. Standardizing permissions on Entra security groups keeps one control model as AI tools multiply.
Enterprise Claude reaches company-wide readiness with security and audit built in
From Anthropic announcements · self-serve Enterprise in Feb 2026 · wider Compliance API coverage in Aug 2026
Security has been the main reason companies hesitate on generative AI. This year Claude Enterprise added the controls that let a company own its accounts, data and logs, and it can now be bought without a sales call.
Self-serve Claude Enterprise (Feb 12, 2026)Buy online and set up SSO, domain capture and SCIM provisioning in minutes. Claude, Claude Code and Cowork for the whole organization, with Microsoft 365 and Slack connectors and Excel and PowerPoint sidebars.
Admin and security controls includedAudit logs, Compliance API, organization-wide permissions, custom data retention, usage analytics and spend caps by organization and user. Enterprise content is not used for training by default.
Compliance API extended (Aug 2026)Covers Cowork (desktop, web, mobile) and Claude Code (CLI, desktop) sessions, including prompts, responses and tool calls, for audit and eDiscovery. Claude Code on the web, API use and Bedrock, Vertex AI and Foundry sessions are not covered yet.
CertificationsSOC 2 Type II, ISO 27001, ISO/IEC 42001, and a HIPAA configuration with a BAA. Check scope by plan and cloud route in the Trust Center.
Takeaway · Moving AI use into a place the company controls is now a practical alternative to blocking it. My Korean handout lays out a seven-step rollout and fixes for eleven security concerns.
PPCC 2026: Power Platform moves from app tool to agent operating model
Power Platform Community Conference 2026 · Oct 27–29, MGM Grand, Las Vegas · 9,000+ attendees
The message this year: apps, automation, agents, code and governance now form one operating model. The edge is shifting from building fast to running agents that work safely.
+20%Power Apps daily active users
+20%Flow runs
2x+AI usage in Power Apps
3x+App usage in Copilot Studio
Year over year · 2M+ monthly active makers and 4M+ apps (Microsoft)
Agents become coworkersCopilot Studio agents act under their own accounts in Microsoft 365, start on triggers and remember business processes. The Microsoft Copilot harness runs longer, complex work.
The line between building and coding blursApps in Copilot Studio (public preview), app creation in Copilot Cowork, natural-language canvas and model-driven apps. Developers pick up the same solution in VS Code, GitHub Copilot and Copilot CLI.
AI that understands the businessBusiness applications in Work IQ (public preview), plus Dataverse semantic models, Ask API, Business Skills and MCP Server.
Code under governanceMicrosoft Copilot Managed Runtime hosts code created with Copilot under enterprise governance.
Governance and spend in one placeMicrosoft Agent 365 gives one view of all apps and agents, action-level connector policies and spending controls by user and asset.
Customer stories · Asurion raised touchless invoice processing from under 20% to over 70% with a Copilot Studio agent; the NFL saves 8 to 12 hours a week of data entry and reporting.
Data
Clean up permissions, table relationships and business terms so AI can read them. Oversharing and poor data quality show up directly as agent risk.
Build
Shared ALM for makers and developers; standards for MCP and GitHub Copilot.
Run
An owner, approver and stop rule for every agent, with a regular inventory review.
Cost
Measure credit use against business results; set limits and alerts per user and agent.
Security
Least-privilege data access and actions; check labels, DLP and connector exposure.
Most features are in preview, with no GA dates or usage limits announced. Start small: assess (0–4 weeks), design (1–2 months), pilot two or three high-value processes (2–4 months), then scale (4–9 months).
MBA thesis · 2023.08 · Ajou University Graduate School of Business
The impact of nuclear power expansion on Korea's System Marginal Price, forecast with deep learning
Shows that the generation mix chosen in the national power supply plan moves the wholesale electricity price (SMP). Holding settlement prices at 2021 levels and applying the 10th Basic Plan's generation mix, an LSTM model forecasts the 2030 SMP.
−22.2%
Forecast 2030 SMP versus the 2021 actual. LSTM validation MAPE 8.6%, against 13.8% for RNN.
2030 SMP forecast (orange) vs. 2021 actual (blue), LSTM, normalized values
Best Presentation Award, Korean Society for Quality Management, Fall Conference 2022, for the preceding study. DBpia ↗ · Ajou University news ↗
Follow-up presentation: "Forecasting the effect of government policy on electricity cost with deep learning", Korea Intelligent Information Systems Society (KIISS) Spring Conference, May 26, 2023, Yonsei University, Business Analytics session (with Juyoung Kang, Ajou University). Program ↗
Personal research · Blockchain basics
SHA-256, the hash that keeps a blockchain honest
SHA-256 turns any input into a fixed 256-bit value. It is deterministic, one-way and collision resistant, and a one-letter change flips about half the bits ("hello" vs "hellp": 131 of 256). Each block stores the previous block's hash, so changing an old block breaks every block after it; mining repeats SHA-256 with new nonces until the hash meets the difficulty rule.
SHA-256
Blockchain
Personal research · Blockchain
Reproducing the Bitcoin genesis block hash in Python
A Python version of the proof-of-work described in Satoshi Nakamoto’s 2008 Bitcoin paper. It joins the 80-byte block header (version, previous hash, Merkle root, time, difficulty bits, nonce), applies SHA-256 twice and checks the result against the difficulty rule. Mining is the search that raises the nonce one by one until a hash meets that rule.
With the genesis block’s real nonce (2,083,236,893) it reproduces the official hash 000000000019d6689c085ae165831e934ff763ae46a2a6c172b3f1b60a8ce26f.
Written with Claude and verified against the real genesis block hash.
Python
SHA-256
Proof-of-work
Personal research · Deep learning
How a CNN condenses an image and builds on it
A convolutional network slides small filters over an image to keep features such as edges, then pools the result into a smaller summary, layer after layer: edges, then shapes, then objects. A worked example turns a 6×6 image (36 values) into a 2×2 summary (4 values) that still says "there is a vertical edge in the middle". The final features feed a classifier, or are expanded again to rebuild or generate images. Because one small filter is shared across the whole image and every position can be computed independently, CNNs need far fewer weights and run in parallel on GPUs, which is why GPU-trained CNNs such as AlexNet (2012) launched the deep-learning boom. RNNs and LSTMs, by contrast, must process a sequence step by step, a limit that Transformers later removed.
CNN
Deep learning
NumPy
Personal research · Generative AI
How a Transformer reads a whole sentence at once
Instead of reading word by word like an RNN, a Transformer lets every token look at every other token at the same time. Each token produces a Query, Key and Value; Query·Key scores are turned into weights with softmax, and the Values are mixed by those weights. A worked example shows "먹었다 (ate)" attending 25% to the subject, 25% to the object and 50% to itself. Because all pairs are one matrix multiplication, it runs in parallel on GPUs, and stacking many such layers and predicting the next token is how large language models generate text.
Transformer
Self-attention
LLM
CareerEducation · Certification
Career, education and certification
Career
2010.05 – present
POSCO International · Senior Manager
Currently in the DX Planning Group, DX Innovation Office. Company-wide DX planning and delivery; in-house Power Platform expert and instructor. ERP PMO for the Poland subsidiary; ERP selection and PMO (issue management) for the India subsidiary.
Employed since May 2010. The company was POSCO Energy until it merged into POSCO International in January 2023; the 2010–2022 roles below are from that period.
2020.09 – 2022.12
POSCO Human Resources Development Center · seconded instructor
Seconded from POSCO Energy. Taught Python data analysis to about 1,600 learners.
2019.11 – 2020.08
Korea Fuel Cell · secondment
IT carve-out from POSCO Energy: SAP to seven Douzone iCUBE cloud modules, independent groupware and IT operations, office network and printer setup, and training of its first IT team.
2014.07 – 2019.10
POSCO Energy · IT Planning Group
Led SAP operations across all modules, authorization and license audits, closing issues and change approval. PM for the 2015 cell plant SAP rollout; planned the S/4HANA next-generation ERP.
2012.10 – 2014.07
POSCO Energy · PI Task Force
Joined the task force set up after the process diagnosis. Owned the plant maintenance (PM) and quality (QM) modules in the Oracle-to-SAP ERP transition.
2012.04 – 2012.10
POSCO Energy · Process Diagnosis Team
Gathered data, VOC and VOB and wrote the reports for two company-wide process diagnoses (new investment business; power plant construction, operation and maintenance). The results were presented to the POSCO chairman and led to the PI Task Force.
2010.05 – 2010.11
POSCO Energy · Oracle ERP expansion
Process design, build and stabilization of the quality (QM) module.
2010.05
Joined POSCO Energy
My POSCO DX business unit was merged into POSCO Energy.
2008.01 – 2010.05
POSCO DX
Implemented the final inspection technology and process for power generation modules and BOP (balance of plant) products, checking over the control network and other systems that each finished unit worked correctly, and trained the staff to run it. Also set up standard work processes and supported ISO 9001 certification (2009–2010).
2001.12 – 2007.12
Controls engineering · a hands-on foundation for physical AI
Physical AI, where AI moves robots and equipment, runs on top of real control systems. On LCD polarizer attachment equipment, I handled the control link between the machine and the 6-axis articulated robot that received panels and passed them on, and later set up and programmed dispenser robots. Electrical and control network design, PLC (Mitsubishi MELSEC-Q) and touch screen programming, equipment set-up and commissioning across equipment projects: installation of the 135 MW Coloane-B power plant in Macau (gas and steam turbines, HRSG) with US and European engineers; chemical supply systems for LG Philips LCD fabs; polarizer attachment equipment for Samsung LCD; and dispenser robots for electronics makers. Because I have designed and commissioned those control systems myself, I can connect an AI model to equipment data, control logic and on-site commissioning, and talk with equipment engineers in their own terms.
Community
2026 – present
Korea Microsoft Power Platform Association · Technical Director
Co-founded the association; leads practitioner exchange between companies.