About

  • I build and debug systems that have to work at scale.
  • Right now that means LLM serving — distributed inference, KV cache movement, and where the time actually goes under load.
  • Most of my earlier experience is on Android at Samsung — runtime, boot, performance.
  • MS in Computer Science at USC, graduating May 2027.
  • I like problems where careful measurement changes the answer.

Experience

Amazon
Software Development Engineer Intern (Summer 2026)
  • Shipped a self-service deploy option that stands up a disaggregated vLLM topology on Kubernetes, separate prefill and decode workers behind a router with an RDMA KV side channel, added alongside the existing single-worker path rather than replacing it
  • Root-caused per-request KV-transfer latency swinging from 12ms to 800ms by reading vLLM and NIXL internals, showing the driver was descriptor count rather than payload size
  • Built the benchmark and analysis pipeline behind 100+ runs, parsing raw Prometheus histograms into one traceable report so every reported number could be traced back to its source
duration12w
users~200
driverdescriptor count
stackvLLM · NIXL · K8s
Samsung Electronics
Software Engineer (2020 – 2025)
  • Cut application startup latency by 15% in Android Runtime, the C++ and Java execution runtime shipping on millions of devices, by profiling the ahead-of-time compilation pipeline and reworking it around pre-execution profiling, prioritized background compilation, and artifact recovery
  • Resolved 20+ release-blocking runtime and framework issues across organizational boundaries with Google and Qualcomm, shipping upstream platform fixes; triaged and reproduced critical failures from 1,000+ reported cases
  • Worked on boot and recovery paths: separating platform resets from kernel resets, watchdog handling, and regenerating runtime artifacts to pull devices out of reboot loops, shipped across multiple Android releases with 95%+ recovery
  • Three issued patents from this work, all deployed in production: pre-execution profiling, reboot-loop recovery, and usage-driven package compilation
duration4y 10mo
startup15% faster
patents3 issued
blockers20+ upstream
Continental
Software Engineer (2018 – 2020)
  • Embedded platform components for ARM-based automotive systems
  • In-vehicle dashboard platforms in production vehicles
  • Performance-critical and resource-constrained environments
duration2y 7mo
adoption200+ engineers
manual QA30% less

Research

LLM Evaluation — CARE
  • Second author on a paper evaluating how well LLMs simulate community reactions to news
  • Built the revision's experiment pipeline, generating and tone-labeling 36K records across six model and prompt conditions
  • Ran the inter-annotator agreement analysis, comparing candidate metrics and implementing the one suited to ordinal, heavily skewed labels
  • USC ISI, with Nuan Wen and Prof. Xuezhe Ma
rolesecond author
statusmanuscript in revision
scale207 communities
  • LLM Evaluation
  • Batch Inference
  • Annotation Agreement
Difficulty-Aware Routing Between Small and Large LLMs
  • Comparing routing strategies that send easy questions to a small model and hard ones to a large model
  • Measuring whether the cost savings reported in the literature hold up in actual GPU serving
  • Course project in progress, CSCI 566 at USC
modelsQwen3 1.7B / 8B-AWQ
datasetsGSM8K · MMLU
measuresserving cost
  • vLLM
  • Serving Cost
  • Model Routing

Patents

Electronic device for compiling files on basis of designated events, and method therefor
  • Detects and repairs damaged ahead-of-time compilation artifacts, so a device can recover itself instead of falling into a reboot loop
  • US 19/030,679 · 2025 · with K. Jeong, H. Kim, H. Lee
Electronic device for obtaining information used to compile application, and method thereof
  • Pre-launches an application on a virtual display after install, keeping a warm process so the first real launch is faster
  • US 2025/0138792 A1 · 2025 · with K. Jeong, S. Lonchakov, I. Titarenko, H. Kim
Electronic device and method for compiling packages on basis of order obtained by interaction
  • Orders background compilation by how the user actually uses each package, so the apps that matter are optimized first
  • US 18/972,350 · 2025 · with K. Jeong, H. Kim, S. Lee

All three come out of the Android Runtime work at Samsung and are deployed in production. Google Scholar

Education

🎓
University of Southern California, M.S. in Computer Science (AI/ML focused) (2025 – Present)

Projects

Time Slicer
  • Daily check-off tasks and weekly goals (target minutes)
  • Log time in blocks (+15m / +30m / +60m), progress bars and weekly bar chart
  • Next.js, React, TypeScript, Tailwind
  • Live: https://timeslicer-rose.vercel.app/
  • Next.js
  • React
  • TypeScript
  • Tailwind
Echo Slice
  • Slice long-form content (articles, videos, notes) into chunks and review flows
  • Optional AI to summarize or structure material
  • Web app with backend for persistence and sync. In active development
  • Live: https://echoslicefront.vercel.app/
  • Web
  • Backend
  • AI