Seunghyun Ryu · 류승현
I work on grounded RAG and faithfulness/abstention eval — getting AI to cite a real source, or honestly say "i don't know," on the high-stakes, slow-to-read documents that bottleneck teams. Proof is public and IP-clean. Deep & reliable over fast.
Public proof
A faithfulness + justified-abstention benchmark for regulated docs: a correct refusal is a first-class pass, scored only for the right reason. I ran three systems, including a real LLM — and showed the default RAG metric returns NaN for a correct refusal. Offline; every gold label adversarially verified against the corpus.
A documentation Q&A system with one rule: cite a real source, or honestly refuse — and refusal scores as a first-class eval. Before publishing, an adversarial review caught it confidently answering questions it had no source for; I found the root cause and fixed it. 11/11 eval cases pass, 20 offline tests, no API key.
A Claude Code / Codex companion that measures the one thing nothing else does: how you ask — and how that's changing. A recurring, local check that ends with what to fix next — skills to build from the work you keep re-explaining — not a one-time vanity card. "Nothing leaves your machine" is verifiable in the code.
How I work
Claims are made only from retrieved, cited sources — never from the words a question happens to share. Absence of evidence is reported, not papered over.
Anything irreversible — sending, modifying infra, exposing data — waits for a person to review and approve. The system drafts; the human decides.
Proof lives on public, synthetic data. Company systems and customer data stay private. The method is the point — so the evidence stays reproducible.
Experience
Toolbox