2024.07 – 2025.06

KMA Road Weather Information System — build

Built road-weather aggregation batches and statistics screens full stack, improving SQL, service-layer responsibilities, and shared JSP structure.

Owned scopeRequirements · backend and frontend implementation · testing · release support

Key outcomes

• Reduced an Oracle statistics query from 15–20 seconds to about one second
• Minute-level source data → 10-minute, hourly, and daily statistics
• Shared JSP table, chart, search, and date handling
• Vanilla JavaScript screen-state helper

Upfront design

• Time-series aggregation — separate collection and reference times; explicit windows
• Service layer — centralized processing and separated layer responsibilities
• Screen state — one flow for search, reset, refetch, and chart updates

Troubleshooting

• 15–20s query → execution-plan analysis and SQL restructuring → about 1s
• Method overloading and mixed responsibilities → service-layer restructuring
• Repeated screen handling → shared JSP structure and state helper

AI document writing · state modeling · process debugging

brieFFlow

An AI workspace for grounded press-release generation and quality evaluation.This case focuses on the debugger built to make that writing workflow trustworthy and easier to improve.

An AI writing tool that models the path from rough notes to drafted, reviewed, and revised copy as traceable state transitions

Owned scopeAI-assisted note structuring, draft generation, draft review, selected revision, state transitions, per-stage checkpoints, and a process debugger

Key outcomes

• Implemented five AI-assisted stages from note structuring through drafting, review, and selected revision
• Modeled writing around state transitions so the current state and debugging point stay explicit
• Used RAG to retrieve and link team-document evidence so critical facts in drafts could be checked
• Added a checkpoint debugger to inspect failures and compare process changes

Upfront design

• State model — defined document initialization, note structuring, draft generation, review, and selected revision
• State transitions — separated each stage's input, output, and conditions for moving forward
• Human checkpoints — required confirmation of structured notes, generated drafts, and selected review feedback
• Execution responsibility — let the domain process, rather than UI call order, own current state and valid transitions

Troubleshooting

• Final output hid the failure point → stored per-stage checkpoints and transition results
• Drafts could drop critical facts → retrieved and linked team-document evidence with RAG, then checked preservation of numbers, dates, and quotes
• Process changes were difficult to assess → compared original and changed attempts in the debugger

AI workflow analysis · safety rules · change validation

AI Process Console

A console that finds where and why AI work failed, then validates changes without affecting production

Owned scopeWorkflow visualization, error trends, safety and quality checks, before-and-after testing, and PostHog and LangSmith collection

Key outcomes

• Rebuilt execution records as a flow of work stages and routes
• Found periods where safety-rule warnings or stops repeatedly occurred
• Compared current settings with proposed changes without touching production
• Prepared PostHog and LangSmith collection for future unified site analysis

Upfront design

• DDD — defined responsibilities for processes, attempts, stages, routes, and requirements
• Event-driven design — rebuilt overall progress from records emitted by each stage
• Hexagonal architecture — separated core business rules from UI and storage technology
• Responsibility split — source projects execute; the console analyzes and requests tests

Troubleshooting

• Branches and retries duplicated routes → kept every run record but displayed each route once
• Mixed success criteria distorted results → separated reached, evaluated, and completed populations
• Reproducing safety-rule failures required live changes → tested copies while preserving the original
• External metrics looked causal → kept them as supporting signals, not failure verdicts