Ethan Justice
Software engineer. Interested in how things work.
Finishing a CS degree at Michigan, starting at Google YouTube in June. This is where I put things I’ve built, things I’m thinking about, and things I learned later than I should have.
writing
- Q1 2026
Finishing Michigan, landed Google, and learned that technical work needs the right context.
- Why I Rebuilt This Site
I rebuilt this site because the old version was a dressed up resume and this one is trying to be more.
selected work
Software Engineer
Backend infrastructure and LLM orchestration for YouTube Ask. C++ and Go.
Machine Learning Researcher
Focused on KV-cache management in LLM inference systems, working within the vLLM and LMCache codebases. The goal was improving prefill speed by implementing bidirectional loading with overlapping IO and compute. A100s and H100s, profiled with Nvidia Nsight.
AI Engineering Intern
One of the first engineers at an agentic memory platform. Built the multimodal pipeline, a semantic caching layer that cut API response time by 99%, and the LLM-as-a-Judge testing suite to verify memory usage.
Heterogeneous Multi Agent Debate
Capstone research project on Heterogeneous Multi Agent Debate for LLM inference. Confidence based gating to avoid redundant debate, cutting total FLOPs by 40%. Identified syntactic determinism as a class of failure modes that makes the approach unsuitable for syntax heavy fields, but reduces hallucinations in factual retrieval.
about
I started coding in middle school writing control software for a robotics team. Went 30 miles south to the University of Michigan for CS.
Graduated in three years. Interned at UWM and Little Caesars doing backend work, then joined PersistOS as one of the first engineers. Spent my last semester at Michigan doing ML inference research on KV-cache management in vLLM and LMCache.
Starting at Google YouTube in June. I don't know what comes after that.
education
University of Michigan