Jeng-Yue (Buffett) Liu
Website

Experience

Work Experience

Company logo

LinkedIn Corporation

Machine Learning Engineer Intern
Focus: Multi-agent systems for developer productivity
Location: Sunnyvale, CA
Dates: May. 2026 - Aug. 2026
  • Engineered a multi-agent tracing and profiling system supporting 60k+ daily PRs, analyzing failure patterns, latency, token usage, and sub-agent parallelization bottlenecks to power a reflective prompt evolution framework that drives continuous system self-evolution.
  • Prototyped a reproducible benchmark assessing multi-agent alignment with historical human reviewer perspectives, using deterministic unit-test execution over LLM-as-a-judge to validate true error detection through fail-then-pass logic.
Company logo

Neutone Inc.

Research & Development Intern
Focus: Real-time neural tone morphing
Location: Tokyo, Japan (Remote)
Dates: Dec. 2025 - Apr. 2026
  • Ported the in-house real-time tone-morphing plugin to a SlowFast training pipeline, mitigating low-buffer granular artifacts and degraded timbre transfer to improve out-of-distribution reliability while preserving low-latency real-time inference.
Company logo

Academia Sinica

Machine Learning Research Intern
Mentor: Prof. Li Su
Dates: Jul. 2024 - Aug. 2025
  • Outperformed state-of-the-art models on 2,500 hours of audio with a 47.3% reduction in multi-scale STFT loss by proposing a factorized codec with attribute-specific auxiliary tasks and information perturbation, enabling controllable disentanglement in style transfer.
  • Achieved 86% k-NN top-1 similarity across 75k+ Beatport segments with a zero-shot timbre encoder built on MoCo-v2 and a Swin Transformer, using sequence perturbation and temporal augmentation for timbre-invariant representation learning.
Company logo

Quid Inc.

Machine Learning Engineering Intern
Mentor: Larrick Chen
Team: Discover Analysis
Dates: Dec. 2024 - Jun. 2025
  • Cut manual prompt tuning by 10+ hours per week by optimizing search-result similarity ranking and match scoring with DSPy under Chain-of-Thought and MIPROv2, and automating summary and title generation via an LLM-based assessment module.
  • Improved emerging-hashtag Precision@50 by 18% by combining BOCPD change-point signals with creator-conditioned engagement features in a LightGBM classifier, enabling early detection of volatile trends across 1M+ creators.

Research Experience

Music and Audio Computing Lab, Academia Sinica

Dates: Jan. 2024 - Aug. 2025
  • Proposed a novel end-to-end factorized codec learning framework for timbre/style transfer models with information perturbation and supervision, achieving enhanced timbre-content-ADSR disentanglement for controllable synthesizer preset conversion and surpassing state-of-the-art synthesizer timbre transfer baselines with a multi-resolution STFT loss from 5.69 to 2.22. [GitHub]
  • Developed an audio-query music source separation system using band-split Mamba2 with hypernetwork conditioning, enhancing timbre conditioning and boosting instrument-specific SNR by 7%.

Productivity Optimization Lab, NTU

Dates: Dec. 2023 - Dec. 2024
  • Collaborated with a team of 6 to develop a GraphRAG-based news content analysis tool, leveraging LLMs for insight extraction from large datasets and reducing manual effort in social science research. [GitHub]
  • Applied NLP techniques such as LDA and NMF to analyze attitude shifts surrounding the 2021 Atlanta spa shootings.

Geospatial Computing Lab, NTU

Dates: Jan. 2022 - Jun. 2024
  • [B.S. Thesis] Trip-purpose-based methods for predicting human mobility’s next location
  • Developed multimodal spatio-temporal models with a trip-purpose approach, achieving 80% accuracy, and designed a data pipeline to integrate mobility data, Google Maps polygons, and remote sensing datasets for spatial analysis.
  • Reached IMV contest semifinals by building a digital transaction platform with TypeScript (React), Node.js, and MongoDB, and developing a Selenium web crawler for real-time vegetable prices to optimize fertilizer use. [GitHub]
  • Simulated 3D crowd and vehicle flows in NetLogo and Python for safer Taipei Dome evacuations, informing exit planning.

Open Source Contribution

Public contributions from @buffett0323, mostly across the LLM inference, serving, and kernel ecosystem.

17
Pull Requests
9
Merged
5
Upstream Projects
168k+
Combined Stars
linkedin/Liger-Kernel★ 6.6k

Efficient Triton kernels for LLM training

  • merged#1390Add Liger Kernel support for Muse Glimmer
  • merged#1326[Megatron] Add SwiGLU integration
  • merged#1358, #1362, #1364, #1366Pinned fp64 scalar kernel parameters to fp32 across cross-entropy, RMSNorm, and GRPO loss kernels
  • open#1396[Bug] FLCE Triton fallback
huggingface/diffusers★ 34k

State-of-the-art diffusion models for image, video, and audio

  • merged#14119Add Stable Audio 3 pipeline
  • merged#14182Fix path-traversal / arbitrary out-of-directory file read vulnerability
  • open#14126Add from_single_file() support for Ideogram4 and Krea2 transformers
sgl-project/sglang-omni★ 1.1k

Serving framework for audio and omni-modal models (TTS, ASR)

  • merged#1420[dots.tts] Fix non-deterministic reference encoding; add parity test
  • open#1459[dots.tts] Measure codec lock contention with GPU span timing
  • open#1434[dots.tts] Add per-call-site contention profiling for the shared codec
  • open#1166Fix code2wav initial chunk size
vllm-project/vllm★ 91k

High-throughput, memory-efficient LLM inference and serving engine

  • open#48579[BugFix] Drafter attention backend auto-selection picks FlashAttention incorrectly
sgl-project/sglang★ 36k

High-performance serving framework for large language models

  • open#31602Fix malformed Harmony output for gpt-oss with reasoning and JSON mode
  • open#30189[LoRA] Fix crash loading GDN in_proj_qkv / in_proj_z adapters

Teaching Assistant Experience

Machine Learning, NTU

Dates: Jan. 2023 - Jan. 2024 (2 semesters)
  • Designed assignments & projects, and led TA sessions in English to support students with problem-solving and queries.
  • Conceived and led the final project, originating from my idea. Utilized Generative Adversarial Networks (GANs) to generate a noisier dataset based on the original, increasing task difficulty in a student final project, then applied machine learning models to establish baselines.

Open Review

ICASSP 2026
Conference: IEEE International Conference on Acoustics, Speech and Signal Processing
Dates: 2026
Description:
  • Reviewing submissions for ICASSP 2026.
NeurIPS AI Music 2025
Conference: NeurIPS Workshop on AI for Music
Dates: 2025
Description:
  • Reviewing submissions for the NeurIPS Workshop on AI for Music 2025.

Skills

Languages

  • Python
  • C
  • C++
  • CUDA
  • Java
  • JavaScript
  • TypeScript
  • R
  • Swift
  • SQL
  • Shell

ML / Frameworks

  • PyTorch
  • Triton
  • NumPy
  • Librosa
  • Hugging Face
  • Diffusers
  • LangChain
  • DSPy
  • JAX
  • vLLM
  • SGLang
  • React
  • Next.js
  • FastAPI
  • Flask

Infra / DevOps

  • Docker
  • Kubernetes
  • Helm
  • Argo CD
  • Linux
  • Google Cloud Platform
  • GitHub Actions
  • Postman

Data / Tools

  • PostgreSQL
  • MySQL
  • MongoDB
  • Supabase
  • Qdrant
  • Neo4j
  • Apache Kafka
  • Prometheus
  • Grafana