April Yang

Hi, I'm April! I'm a research engineer at NVIDIA, working on vision language models training and evaluations, multimodal understanding, and agentic data search & curation flywheels, with a domain focus in Physical AI.

Outside work, I enjoy photography, traveling, driving with music, and spending time with my three cats.

CV / Scholar / LinkedIn / GitHub

Expertise

Data Flywheels

Multimodal search and data flywheels for physical AI

Building multimodal embedding-based retrieval and curation system over petabyte-scale video datalake, supporting data flywheels that powers end to end training loops.

VLM Post-training

Training and evaluating VLMs for AV scene understanding

Post-training VLMs for driving scene understanding, building holistic evaluation and benchmarking covering caption quality, temporal reasoning, embedding quality, and retrieval behavior, deriving insights for model selection and identifying failure modes.

Agentic Loop

Agent loops that curate long-tail driving scenarios

Building an agentic search-refinement loop that iteratively search, labels, refines, to curate datasets, with an evaluation harness that keeps the agents measurable, reproducible, and self-improving.

Publications

Overview of the parallel prompting pipeline: ego trajectory tags are injected into parallel prompts, a VLM writes section captions, and a merger consolidates them

Scalable Parallel Prompting for Complex AV Video Captioning

April Yang*, Roberto Amoroso*, Nikita Durasov, Devansh Bisla, Sandipan Kundu, Elmar Haussmann, Ruchi Bhargava, Maying Shen, Nadine Chang, Jose M. Alvarez

CVPR Workshops (DataMFM), 2026

paper /

A scalable video captioning framework based on small VLMs that produces diverse intermediate captions, which are consolidated into a comprehensive unified description for driving video retrieval.

* Equal contribution

@inproceedings{yang2026scalable,
  title={Scalable Parallel Prompting for Complex AV Video Captioning},
  author={Yang, April and Amoroso, Roberto and Durasov, Nikita and Bisla, Devansh and Kundu, Sandipan and Haussmann, Elmar and Bhargava, Ruchi and Shen, Maying and Chang, Nadine and Alvarez, Jose M.},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  year={2026}
}

SIL-Wheel mascot: a car holding a pen, from the open-source repository

SIL-Wheel: A Multi-Modal Search and Curation Platform for Physical AI Systems

NVIDIA (core contributor)

Whitepaper, 2026

code / paper coming soon

An open-source system for large-scale video data curation that fuses complementary retrieval signals over tens of millions of clips, closing the loop from model failure discovery to training and evaluation.

FlexLLM system overview: inference and finetuning tokens co-served on shared GPUs

FlexLLM: Token-Level Co-Serving of LLM Inference and Finetuning with SLO Guarantees

Gabriele Oliaro, Xupeng Miao, Xinhao Cheng, Vineeth Kada, Mengdi Wu, Ruohan Gao, Yingyi Huang, Remi Delacourt, April Yang, Yingcheng Wang, Colin Unger, Zhihao Jia

USENIX NSDI, 2026

paper / arXiv /

The first system to co-serve LLM inference and parameter-efficient finetuning on shared GPUs. Token-level fusion lets finetuning soak up idle capacity while inference keeps its latency SLOs.

@inproceedings{oliaro2026flexllm,
  author = {Gabriele Oliaro and Xupeng Miao and Xinhao Cheng and Vineeth Kada and Mengdi Wu and Ruohan Gao and Yingyi Huang and Remi Delacourt and April Yang and Yingcheng Wang and Colin Unger and Zhihao Jia},
  title = {{FlexLLM}: {Token-Level} {Co-Serving} of {LLM} Inference and Finetuning with {SLO} Guarantees},
  booktitle = {23rd USENIX Symposium on Networked Systems Design and Implementation (NSDI 26)},
  year = {2026},
  isbn = {978-1-939133-54-0},
  address = {Renton, WA},
  pages = {1359--1379},
  url = {https://www.usenix.org/conference/nsdi26/presentation/oliaro},
  publisher = {USENIX Association},
  month = may
}

AdaServe compared with continuous batching: per-request speculative decoding meets each latency target

AdaServe: Accelerating Multi-SLO LLM Serving with SLO-Customized Speculative Decoding

Zikun Li, Zhuofu Chen, Remi Delacourt, Gabriele Oliaro, Zeyu Wang, Qinghan Chen, Shuhuai Lin, April Yang, Zhihao Zhang, Zhuoming Chen, Yi-Hsiang Lai, Xinhao Cheng, Xupeng Miao, Zhihao Jia

EuroSys, 2026

arXiv /

An LLM serving system for requests with different latency targets. SLO-customized speculative decoding builds a speculation tree per request, meeting heterogeneous SLOs while maximizing throughput.

@inproceedings{li2026adaserve,
  title={AdaServe: Accelerating Multi-SLO LLM Serving with SLO-Customized Speculative Decoding},
  author={Li, Zikun and Chen, Zhuofu and Delacourt, Remi and Oliaro, Gabriele and Wang, Zeyu and Chen, Qinghan and Lin, Shuhuai and Yang, April and Zhang, Zhihao and Chen, Zhuoming and Lai, Yi-Hsiang and Cheng, Xinhao and Miao, Xupeng and Jia, Zhihao},
  booktitle={Proceedings of the 21st European Conference on Computer Systems (EuroSys)},
  pages={36--54},
  year={2026}
}

Full publication list on Google Scholar