From user interfaces to operations

• Frontend — Chat, Agent, Generative UI, WebSocket, admin back office, file previews, and user-facing visualization of RAG processing and chunking/embedding status. Also developed a VOC Agent that creates Jira tickets from Generative UI forms.
• Backend / API — Python/FastAPI and Go experience covering metering queries, usage limits, resource permissions, and sharing APIs. Go Gateway ownership is limited to the features I directly implemented.
• Enterprise — Multi-organization RBAC boundaries for read / execute / write / share, Agent sharing approval, revocation, and execution reauthorization.
• Data / Operation — Query APIs integrated with shared metering ingestion, daily aggregation, and price/FX infrastructure; Docker, testing, deployment, Smoke Tests, and operational automation.

Within six months of joining, I worked with the team on the first release for approximately 10,000 people across 13 affiliates. This does not imply ownership of the entire platform or shared infrastructure. RAG work means UX and integration for processing status, not development of chunking algorithms, embedding models, or a Vector DB.

Technical evidence retained

Separate case studies retain implementation, verification, and limitations for usage/cost metering, multi-organization Agent sharing, in-app file previews, Generative UI JSON recovery, WebSocket reliability, and Docker image optimization.

From inspection requests to reports and history

• Frontend — React, TypeScript, and TanStack Query for inspection creation, file/repository input, progress, traffic-light outcomes, Risk Graph, improvement recommendations, and history.
• Backend — Node.js, TypeScript, and Clean Architecture connecting file/ZIP inputs, inspection of public GitHub/GitLab repositories, asynchronous inspection state, results/history APIs, and JSON/PDF reports.
• Evaluation — Four task-specific quality presets, eight attack personas, five defense personas, and ten core risk scenarios.

Separate deterministic decisions from LLM exploration

The LLM is restricted to extending risk scenarios that ordinary tests may miss. External models receive only minimized, de-identified information, not the complete original inputs.

Human approval, sensitive-data masking, audit logs, duplicate-execution prevention, untrusted-input isolation, and output-path restrictions are product controls. Safety, permissions, auditability, and reproducibility matter alongside Agent performance.

This was development work in a Hansol Group AX project. Validation results are limited to the defined inspection scope, not a guarantee that every risk will be detected.

Apply a model to an enterprise system

The Agent interprets the request and selects purpose-defined MCP tools. It queries production, quality/defects, equipment, and LOT history; compares the current period with the previous equivalent period; and queries additional data for unusual changes before producing a summary and HTML report.

This is experience applying an on-premise model to an Agent and MCP—not training or building the model, or developing the entire MES.

Separate observed facts from interpretations

• Restrict numerical claims to MCP query results.
• Separate observed facts from possible causes and bound query ranges.
• Verify MCP tool error retries and defenses against requests to fabricate production figures.
• Distinguish model, prompt, MCP, and infrastructure failures during diagnosis.
• Include normal requests, out-of-scope queries, fabricated-data requests, and infrastructure failures in normal, exceptional, and adversarial QA.

Loop engineering connects implementation to testing, review, deployment, and observed failures. Harness engineering gives AI tasks explicit inputs, tool procedures, completion criteria, and verification rules. AI output is checked through actual code and screens.

The five practices below cover development quality, weekly reports, guide/web-document linking, screen-design/test-case support, and Excel processing. Reporting time refers to the recorded generation and validation task, not overall productivity.

Loop engineering connects implementation, verification, observed failure, and revision. Harness engineering provides the inputs, tools, completion criteria, and checking procedures for AI-assisted work.

After defining states and impact, I used AI for implementation and test authoring, then checked unit tests, screen QA, GitLab MR, deployment pipelines, and smoke behavior. Failures became regression tests and updated checking rules. This does not claim every stage runs automatically for every task.

Collected DAU, tokens, top Agents, and model usage from APIs, then checked periods, totals, and week-to-week continuity. AI explained verified changes and unusual items; a person reviewed the final report.

Recorded generation and validation for one weekly report went from approximately 2–3 hours to under 10 minutes. This applies to that task, not every operational activity.

Captured real screens, added annotations/descriptions, and reused them in guides. During the AX Campus migration, I connected scenario contents, internal links, indexes, and image references.

Five guides and 93 actual screen images describe output scale. A separate usage-video task went from ~3 days → 4 hours; this is not a measured saving for the web-document migration alone.

Connected screen IDs, descriptions, and images to support semi-automated HWP assembly. Reviewed document order, duplicates, and omissions, then used AI for test-case drafting and executed actual functions.

Completed the screen specification and final HWP while checking screen consistency and wording personally. Authoring assistance is separate from execution; this does not claim full test automation.

Used AI to assist code and explanation authoring for recurring Excel processing, validation, and reports. Actual calculations and outputs run in code; I checked values, omissions, columns, and formatting.

This scope draws on PTKOREA DNT validation/Excel reporting and internal administrative tasks. No additional unsupported time-saving rate or autonomous-analysis outcome is claimed.