About
I build AI agents for biomedical research and check how far their conclusions can be trusted. My background is bioinformatics and clinical genomics: multi-omics analysis, Phase 1 to 3 clinical trial data, and reports and services that go to clinicians and pharma clients. That work now runs through LLMs and agents. I work at Cytogen as Chief Researcher in the Diagnosis Division.
2026 highlights
- 1st place, NVIDIA Agentic AI Hackathon 2026: I led FlyGate, a five-person team that built an agent workflow on the NVIDIA stack, from protein-drug docking to post-market adverse event review. The critic agent rejected all 16 planted overclaims and flagged none of 17 clean cases.
- ML4H 2026 Findings: first and corresponding author of a paper on a reliability map for per-gene RNA velocity parameters. The same study is an oral talk at BIOINFO/GIW ISCB-Asia 2026 on November 18.
My part on the NVIDIA AI stack
For FlyGate I built the writer and critic agent harness on NeMo Agent Toolkit and Nemotron, wired DiffDock docking in through BioNeMo NIM, and wrote the OpenShell policy that limits which files and hosts the agents can reach. I earned NVIDIA Certified Associate: AI Model & Application Development in November 2026.
Current work
- Built a model that assists cell image review, deployed it in house, and documented where it applies and where it does not.
- Designed, built, and now run four internal data tools together with the portal that ties them together.
- Leading two Pseudo Lab projects: AutoBioX, a multi-agent research harness, and Falsify, a benchmark that measures how well AI reviewers catch planted flaws in papers.
- Presenting VeriVar, an LLM variant interpretation pipeline with citation checking, as a poster at K-BioX Global Bio Innovation Summit 2026 in Chuncheon on November 14.
- Speaking at AWS Community Day 2026 on November 22 about multi-agent workflows for biomedical research.
What this blog covers
How the tools and pipelines I build actually came together, how far models and benchmarks hold up once they are checked, what designing AI agent harnesses has taught me, and what I find when I re-measure the numbers a paper reports. Most posts go up in both Korean and English.
Links and contact
- One-page profile: kakyungkim.github.io
- English CV: KaKyung-Kim-CV.pdf
- Korean CV: KaKyung-Kim-CV_KR.pdf
- GitHub: kakyungkim
- LinkedIn: kakyungkim
- Google Scholar: profile
- Email: kakyung.kim@gmail.com