Switch language / 언어 전환
Decision case studies, told as judgment / 의사결정 케이스 스터디
Product Planning × AI-Augmented Engineering
My background is in operations, business planning, and data analysis. I turn work problems into product flows, build with AI, and use implementation and experiment evidence to decide what to change or stop.
|
1 minute · ad_ko.mp4 ad_ko.mp4 |
3 minutes · Intuitive
|
More 1-minute videos
|
English ad_en.mp4 |
1min_v3_master.mp4 1min_v3_master.mp4 |
More 3-minute videos
|
Animation
|
Museum
|
|
Standard · Version 1
|
Standard · Version 2
|
| Case | Problem and choice | Evidence boundary |
|---|---|---|
| Customer-center staffing | Used Q90 and a conservative option to plan for high-load days hidden by averages | Three fewer staff than the previous year, saving ₩6 million; post-adjustment connection rate exceeded the 90% target |
| Asset generation and review | Let reviewers compare, regenerate and review outputs, then continue unfinished work | 16,052 reviewed assets deployed and license terminated; generation API spend ≈40% of the former annual license fee |
| NLP category matching | Connected natural language to existing categories and updated embeddings only for changed data | Since March 2026, used by production agencies for v비즈링 and v프로필 across three mobile carriers; owned the automatic-production module |
| Internal automation | Connected repetitive reporting and collection to web tasks with human exception handling | Monthly reports for four services every month since March 2025: two days of work by an experienced planner generated in under half a day |
AI Writer System · In development
A writing workspace for long-form creators to find earlier settings and events while drafting. Consistency tracking is the product hypothesis, so memory retrieval and review sit alongside generation.
Draft → check the source behind a memory suggestion → approve or reject → retrieve it during later writing. Generated prose arrives in a side pad and changes the manuscript only when adopted. Drafting, generation, memory review and retrieval are connected in one local workspace. I use it for my own writing, dogfooding the product and improving the analysis display and review workflow from issues encountered during use.
Experience and implementation evidence
Long-Form AI Video Production System · Company project, quality under evaluation
Designed an experience where requesters work with video intent and scenarios instead of generation workflows. Reference-image review feeds the generation plan; execution recipes stay internal.
In development within a CEO-led task force, focused on B2B advertising. I perform most testing; operational adoption is a later stage.
My project contribution is 85%. I own service planning, architecture, the full backend and management of two working developers; frontend development is handled separately.
Generation and master assembly run locally. Custom-node reference features support long-form continuity and voice consistency to a degree; differences in output quality across models remain the focus of evaluation. Source is private.
- Verified RAG — An MVP for foundational AI-service capabilities in the CEO-led task force. My project contribution is 90%, and I currently test it during development.
- Assessment Spec Harness — A supporting PoC checking assignment/rubric mismatches, with deterministic reproducibility checked on synthetic examples.
- Agent Memory System — Explored shared memory across AI clients. The canonical-record/search-cache separation carries into AI Writer.
- Logo Workbench · Harness IR — Research into segmentation failures and structured extraction, distinct from validated product outcomes.
- Constraint-search research series — HW-WFC, Circle-WFC, T-WFC. Recorded tested successes and failure conditions. Corridor generation remains a follow-up hypothesis for Circle-WFC.
- Q-PSA — A separate perturbation experiment inspired by WFC. Stopped after layer removal raised PPL 3.65× versus 1.05× for the baseline, with scoring roughly 1,300× slower.
Start with the problem and existing task, choose the experience and scope, then build with AI and preserve regression evidence. Working implementation and user benefit require different evidence. Connect each product decision to the relevant evidence.
Decisions and technical evidence — PROJECTS.md
Email: kdtyohan@gmail.com
LinkedIn: entangelk