Resume | Aki's Blog

Education

University of Rochester · Rochester, NY

B.S. in Computer Science and B.S. in Applied Mathematics, expected December 2026
Dean’s List: Fall 2023 – Spring 2025, Spring 2026
Full course list: Courses

ML Systems Research, University of Rochester (Prof. Sreepathi Pai) · May 2026 – Present

Study whether a GPU kernel autotuner can fix low-impact tuning parameters to cut tuning cost while keeping the quality of the final kernel.

Research Assistant, HCI & Spatial Computing Security, BEAR Lab, University of Rochester · Sep 2025 – Dec 2025

Analyzed sensor-driven attack vectors in AR/VR, including side-channel leaks and redirected-walking manipulation.
Prototyped interaction dynamics in MuJoCo to explore security issues such as sensor drift in virtual environments.

Artisk Inc. · Software Engineer Intern · San Francisco, CA (Remote) · May 2025 – Feb 2026

Implemented backend features for a deployed internal log management tool with FastAPI, SQLite, and Amazon S3: date-based log rotation, compressed archives, and log search and download.
Replaced backend-proxied downloads with S3 pre-signed URLs, and fetched archived logs on demand for search with automatic cleanup of temporary files.
Built a multi-step onboarding questionnaire that stores answers in MongoDB, and AI name suggestions for new collections using debounced OpenAI requests with a local fallback.
Built admin pages for user feedback and subscription plans across an Express API, its tests, and React views.

Dentist Journey · QA & Test Engineer Intern · San Diego, CA (Remote) · Sep 2025 – Dec 2025

Tested authentication, AI-generation, subscription, and file-upload APIs; documented inconsistent responses, unenforced usage limits, upload-validation failures, and N+1 queries in bug reports for the founder and engineering team.

LLM Quantization Tax Profiler · Mar 2026 – Present · Write-up

Built a reproducible single-GPU LLM profiler that separates no-hook end-to-end benchmarking from CUDA Event layer diagnosis for FP16 and bitsandbytes FP4 inference on an RTX 4060 Laptop GPU.
Measured a controlled FP4 latency–VRAM trade-off: 47.7% higher batch-1 decode latency with 56.5% lower peak allocated VRAM, then used layer diagnosis and Roofline reasoning to narrow the profiling question.
Implemented a Triton fused FP4 k/v GEMV prototype, validated its correctness across 56 projections, and reported its neutral end-to-end result instead of claiming a speedup.

MoodStream · Jan 2026 – Jun 2026 · GitHub

Real-time danmaku (live comment) sentiment aggregation platform built with Go, Redis, gRPC, and server-sent events. Archived.

Programming: Python, Go, Java, TypeScript, C, SQL, Kotlin
Frameworks and tools: PyTorch, vLLM, Triton, Redis, React, Node.js, FastAPI, Docker, Git, MongoDB, AWS S3, Hugging Face
Languages: English (full professional), Chinese (native), Japanese (conversational)