I build things with AI, usually because something in real work annoyed me enough to make a tool for it.

I'm a Packaging & Product Development Engineer in South Korea 🇰🇷. I did not start as a software developer — AI lowered the barrier enough that ideas I had been carrying for years could finally become working programs.

Some projects became useful tools. Some turned into research. Some mainly proved that an idea was not worth using.

I prefer testing ideas over just describing them.

I build AI-native workflows that turn real work into verifiable deliverables.

Read the English field report · 한국어

How I use ChatGPT Chat for actual presentations, small apps and media post-production, keep bounded rendering/voice stages in suitable runtimes, and close the loop with evidence. Includes controlled story-led 2D video production, the qualitative impact of Astra, current policy caveats, failure modes and tested reference contracts. Reported cadence: about five 90-second videos published daily; practical peak production estimated around eight. These are author-reported figures, not an independently audited benchmark.

The scheduler I wanted for years because I am very good at accepting work and much less reliable at remembering all of it. It grew from an old personal Excel scheduler into an AI-assisted workflow for reviewable tasks, events, and follow-ups.

Started as an Excel sheet for checking whether an assembly had a tolerance problem. I kept adding Gap/Overlap analysis, sensitivity, and design comparison until a small check became a reusable engineering tool.

A local-first quality-data system for turning repeated OQC evidence into accumulated history, patterns, and reviewable candidates for better inspection criteria — without letting incoming data silently rewrite the official standard.

A memory core for remembering people across meetings: what was discussed, what was requested, what was promised, and what should be brought up next time.

A portable local AI workspace for constrained Windows environments where normal desktop AI/development tooling may not be available. Built around explicit user-selected access rather than security bypasses.

Persistent external knowledge architecture for AI projects.

Model is replaceable. Root persists.

I use Chat to deepen and structure ideas, preserve the useful state outside the model, and hand sufficiently concrete plans to coding agents for implementation.

Started by asking how much of its own execution environment a chat AI could perceive and use. That led to local verification and processing tools that turn multi-step machine work into compact evidence for the AI. Reusable tools later moved to ChatGPT Library because the runtime workspace was isolated per Chat.

Started as Burst Fleet — one AI rapidly splitting into many workers and merging back down. A prototype was built, but the token cost was clearly high while the actual speed benefit remained uncertain; image-editing workloads were the one case that looked somewhat faster. The surviving idea became narrower: let one strong AI keep working while safe, predictable background work is prepared early. Experimental and default-off.

Experiments around extending an ordinary chat workflow through external capability surfaces while keeping the boundaries explicit.

A small experiment to see how far an AI workflow could go by itself. A Useless Discovery.

• Where AI reasoning should stop and deterministic software should start
• How to preserve useful project knowledge across sessions and models
• How to verify AI-built software without letting the verifier pretend to own the judgment
• How AI, local tools, memory, and human decisions can fit together without hiding the boundaries

I'm interested in open-source collaboration with people who want to build, test, break, reproduce, or challenge these ideas.

If something here looks useful — or wrong — I'd like to hear about it.

Contact: tequiero7342@gmail.com