Work

I work across four areas. The details are unpacked in videos and papers, so here I've just laid out what's going on, at a glance.

Wholesale & agricultural data

As managing director at Daejeon Jungang Cheonggua Co., Ltd., I work with the settlement records and shipment patterns that pile up there every day. I run a pipeline that automatically collects and organizes eight years of settlement data.

On top of that, I'm building a conversational agricultural-intelligence platform (Wholesale Intelligence). It's about asking and answering — in plain language — questions on price patterns, shipment flows, and trends by item. The goal is to let the data answer the questions I run into on the floor every day.

Two of these you can open and look at yourself. A map of produce flows across the country's wholesale markets shows where goods actually move from and to — on a map rather than in a table (last four months). A farm management diagnostic dashboard lets a farm work through its own operating position.

Details are covered separately in talks or videos. For collaboration, see Contact.

Research & papers

My research is mostly about how to verify AI systems, with one more paper that looks at my day job — the wholesale market. The actual claims and methods are in each paper. How far can you trust what an AI returns, and how should you handle that — that's the part I find most interesting lately. On the AI-verification side there are four papers. The unified one was not accepted at TACL, so I rewrote it to address the reviews and posted it on arXiv (2610.01471); the wholesale-market study is submitted to the Journal of Rural Studies and under review. Separately, there are two papers applying this to medical error detection — one on AI evaluation methodology (in preparation), and one medical application written with a clinical expert (under review at ACL ARR).

2026.03 CCR (preprint) arXiv:2603.12123
2026.03 D-CCR (preprint) arXiv:2603.16244
2026.03 HCCA (preprint) arXiv:2603.21454
2026.05 → 10 XMR (preprint) — When Does a Second Model Help? (not accepted at TACL → rewritten, on arXiv) arXiv:2610.01471
2026.06 Wholesale-market trust premium — day-job research (JRS, under review) submitted · Journal of Rural Studies
in prep Model or Harness? — variables in LLM-as-judge evaluation in preparation · for TMLR
2026.07 Cross-model consensus for medical error detection submitted to ACL ARR · replied to reviews · awaiting outcome · co-authored

The full list is on Google Scholar. For deeper discussion, please email me.

Security research · open source

I look for and report memory-safety bugs in open-source libraries. In 2026 I filed 15 CVE reports, and for 4 of them I sent PRs upstream. Three are still open. The jansson maintainer agreed the bug is real but said neither the PR nor the patch was acceptable — the description was too long and the comments outweighed the code (2026.07). Fair point — I'll keep them short next time. In May I filed my first MITRE cveform directly (Request 2041287, miniz).

2026 brotli — size computation bounds DCHECK PR #1472 · open
2026 snappy — boundary handling PR #237 · open
2026 protobuf-c — integer overflow review PR #800 · open
2026 jansson — memory-safety hardening PR #735 · closed

The CVE count is as of 2026.05.

Content · Sapjil Coding

I run a YouTube channel called "Sapjil Coding." The rule is to share only what I've actually built and used. No overstating, crediting AI when AI helped, and saying so when I don't know.

I also built MCP servers to make Korean public data (agricultural products, wholesale markets, and so on) available to AI agents. Some are registered in the A2A agent ecosystem. There's also a Naver News Scraper on the Apify Store — for automatically collecting Korean news.

In September 2026 I entered the Agents for Humans hackathon with an agent that refuses to guess. A farmer asks "peaches — which market paid more?" and it compares that day's auction records, and it declines to answer when the records are too thin. The demo video and code are public.

In October 2026 I briefly went back to my old research field and made MXene in 3D — an interactive 3D walkthrough, in eight steps, from the parent crystal Ti₃AlC₂ to a single ~1 nm sheet of MXene (Ti₃C₂Tₓ). Built with Claude and three.js; the ten references were checked against their bibliographic records. The source is public.

In the same month I started the MXene Atlas — a map of published MXene experiments, from composition and synthesis to measured properties. Every value on the map carries the sentence it was taken from and a label saying how far it has been checked. As of mid-October 2026 it holds about 3,400 values from about 730 papers. The values were extracted with AI (Claude), and each was checked against its source sentence by two different AI models; expert human checks have only just begun, so treat it as a starting point for reading the papers, not a replacement for them. Corrections are welcome.

→ videos & writing