Software and AI engineer building reliable backend and ML systems.
MSCS student at the University of Southern California and Graduate Researcher at the USC Information Sciences Institute (ISI), in Jonathan May's group, working on language-model evaluation and multi-agent reasoning. Before USC, about four years of software engineering across backend, data, cloud, and AI systems — including technical lead on production LLM systems for three enterprise client projects.
Based in Los Angeles. Looking for Summer 2027 software engineering and AI/ML engineering internships.
English → Japanese translation evaluation. Building evaluation pipelines that combine COMET-QE, pairwise LLM judges, seeded A/B randomization, and blind human evaluation, so that quality comparisons are reproducible instead of anecdotal.
Multi-agent social reasoning. Self-training for language agents using social deduction and negotiation games.
Publication. Co-author of Enhancing AMR Parsing with Group Relative Policy Optimization, XLLM Workshop @ ACL 2025. My contribution: dataset construction and evaluation.
Backend services, data pipelines, and cloud infrastructure, plus applied LLM work — retrieval (RAG), evaluation harnesses, and the unglamorous part of making model output measurable and reproducible.
Python · TypeScript · FastAPI · Next.js · React · PostgreSQL / Supabase ·
LLM evaluation · RAG · Vercel
amr-summarizer-prototype — detects factual errors in AI-generated summaries by parsing both the source article and the summary into AMR graphs and aligning them, using a SMATCH++-style graph matcher with Sentence-BERT sentence selection. FastAPI · amrlib · Penman · NetworkX · React, with a pytest suite. Research prototype behind my BSc thesis.
Closed-source products I designed and built end to end (Next.js, Supabase, Stripe): KASHITE (lending tracker) · YOMU (document Q&A with citations) · Phrasely (English writing feedback). More at kyren.app.
Away from the keyboard: PADI Master Scuba Diver and Rescue Diver, 60 logged dives.