Journey
A record, in order, of what I built through 2026 between the wholesale market and code. Not only the finished results, but the getting-stuck, the fixing, and the relearning.
Four strands that feed each other
They look like separate things, but they actually interlock. A verification method from research checks the data from my day job, that data becomes evidence for a paper, and the whole process gets unpacked in a video.
Wholesale & agricultural data
Tools for the settlement and shipment records piling up daily at Daejeon Jungang Cheonggua Co., Ltd. Where real people and trade are.
Research & AI verification
How much can you trust what an AI returns? Papers are the means; the goal is tools that actually get used.
Content · Sapjil Coding
Only what I've actually built and used, on video. No overstating; if I don't know, I say so.
How I work with AI
Not a tool that only answers, but a partner that looks ahead. More in Approach.
In order
First papers, out in the open
I wrote up a method for cross-verifying code review with several AI models and put it on arXiv. It mattered as the first thing I'd "put out into the world."
Turning the habit of doubting into a system
I started recording "how I worked" each day and reviewing it. So I don't repeat a mistake, I leave a rule behind every time I get stuck.
Extending it to medical code
I tried the same verification method on finding errors in medical coding. Combining several models' outputs caught things a single model missed.
Started working with an agricultural research institute
To go deeper into agricultural data, I began talking with a research institute. The point where the day job meets research.
A second YouTube channel
I opened a new channel to tell the data-side stories separately.
Open-source security, in earnest
I started finding and reporting memory-safety bugs in open-source libraries — PRs to brotli, snappy, protobuf-c, and jansson, plus CVE reports.
"Change the lens and you see more"
With the same code, splitting the review into several perspectives surfaced far more. A moment of confirming something I'd only known in theory.
Opened this site (manus.mn)
I wanted the scattered things in one place. This is the site you're looking at.
Submitted the unified paper to a journal
I pulled the verification work together and submitted it to TACL. It's under review now.
A tool that verifies citations automatically
I built a tool that checks whether the citations in a paper or document actually exist. That's where it stops — a real source may still not support the claim it's cited for, and judging that is a person's job, not the tool's.
Opening it to the world — an English edition
It felt like a waste to keep this site in Korean only, so I opened an English edition with the same content. Now you can read the same story in either language.
"A partner that looks first," in earnest
Without waiting to be told, it looks around and surfaces what I missed every day. How far it's come and where it's stuck, I write down honestly in Approach.
Writing down the working principles
Do the free things on its own; always check before anything that costs money. Look broadly rather than narrowly, and act proactively. I'm putting these standards into words.
The day job became a paper
When a price is set at a wholesale market, it isn't only numbers that move — trust and long-standing relationships move with them. I wrote that up and submitted it to the Journal of Rural Studies; it's under review. The first time the research turned back toward the day job.
Putting the market's flows on a map
Where produce actually moves from and to, across the country's wholesale markets — on a map rather than in a table. The goal was for the flow to be legible even if you don't read numbers. The last four months are open here.
Checking that nothing went missing
We had eight years of records, but had never actually checked whether all of it made it in. Reconciled against the official published figures — no discrepancy. But this only says "our archive is complete," not "the underlying numbers are right" — they come from the same source. That's as far as this check reaches.
A medical paper goes out too
A method for finding errors in clinical records by overlaying the judgments of several AI models. Written with a clinician, submitted to ACL ARR, currently under review.
We took down our own scorecard
I had been describing the citation-checking tool with a performance figure measured months earlier. When I went to measure it again, the measurement no longer reproduced. So I removed that number from the public copy. Better to say only what still holds up when you measure it today than to keep a flattering figure around.
The place the watchdog wasn't watching
I had automated checks so the copy couldn't quietly drift back into overclaiming — but the first screen people actually read sat outside those checks. I widened the scope, then deliberately planted wrong copy to confirm the checks really fire. A green light you can't explain isn't a green light.
Still running today
Making a lot turned out to be easier than keeping what you made alive. These are the things running every day right now.
- Wholesale data pipeline — collecting and organizing eight years of settlement records automatically; the last four months are open as a map
- Agricultural intelligence — asking and answering price and shipment flows in plain language
- AI verification research — the unified paper and the rural-market paper, both under review
- Citation checking tool — tightening the copy and the checks so it claims only what still measures today
- Open-source security — submitted PRs waiting on maintainer review
- Content — turning what I've actually done into video