Ryan Kim

Drove 37% of the gain over baseline on MedARC's NanoPath benchmark with an LLM autoresearch pipeline built around a dual-branch anti-overfitting architecture. Separately built an independent local pathology-model robustness pipeline similar to WAIV's work, surpassing their reported results on 7 of 8 tasks and improving scanner-noise robustness by 34%. Post-trained Qwen3.6-27B with GRPO, cutting reasoning-token usage by 30% without benchmark loss, served locally with vLLM.

DNABERT-based prognostic model for glioblastoma survival via Cox regression on non-coding regulatory DNA mutations from 185 patients; identified age-related survival signals and racial disparities encoded in mutational scores.

Benchmarked 12 pathology foundation models across TCGA and internal cohorts, finding a 2x increase in out-of-distribution tumor-purity error (MAE ~0.10 to ~0.20) and highlighting the resulting clinical deployment risk.

Led engineering for a high-concurrency math competition platform serving 150k+ users across 11 partner organizations, building the React/Node.js authentication, submission, and leaderboard workflows. Raised $20k in funding (Hudson River Trading, Innovate901 First Place).

Built a CPU-only, locally-deployed LLM pipeline (LLaMA 3.1-8B, Qwen3-4B, DSPy/MIPROv2) to classify 370 clinically annotated lung cancer CT reports; co-authored a referral-event analysis abstract accepted at the ACRO 2026 Summit. Applied Cox regression and survival analysis to 230 patients' records, linking percent thymic tissue to NSCLC outcomes.