Senior ML Engineer · Uber AV Labs

Art Jiang

San Francisco Bay Area · robotics · machine learning · GPU systems

Abstract

I build perception and machine learning systems for robots — real-time vision, depth estimation, and inference pipelines running on edge hardware. Currently working on autonomous vehicles at Uber AV Labs, after years of shipping imaging systems for surgical robots. My research spans vision datasets, image restoration, video world models, and applied LLM systems.

EmailGoogle ScholarLinkedIn
Portrait of Art Jiang
SurgiSR 4K: 1080p versus native 4K endoscopic detail
SurgiSR 4K. 1080p (top) vs. native-4K (bottom) endoscopic detail, from the dataset paper.
iMED multi-endoscope 3D reconstruction: RGB, depth, aligned point clouds, geometry, and tool labels
iMED. Multi-endoscope 3D reconstruction — RGB, depth, aligned point clouds, geometry, and tool labels.
Seeing Through Smoke: smoky input, desmoked output, and estimated smoke layer
Seeing Through Smoke. Input with surgical smoke, desmoked output, and the estimated smoke layer.

§ 01

News

  • 2026Joined Uber AV Labs, working on machine learning for autonomous vehicles.
  • 2026New preprint: “Seeing Through Smoke” — real-time surgical desmoking, with the 5,817-pair SeeThroughSmoke stereo benchmark. [link]
  • 2025SurgiSR 4K, our native-4K endoscopy dataset, published in MELBA and presented as a MICCAI 2025 oral. [link]
  • 2025“Beyond Retrieval” — ensembling cross-encoders and GPT rerankers for biomedical QA — published at CLEF 2025 (BioASQ Task 13b). [link]
  • —Demoed the Da Vinci 5 surgical robot at NVIDIA GTC. [photos]
  • 2023Promoted to Senior Machine Learning Engineer at Intuitive Surgical.

§ 02

Papers & Datasets

Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception
Fig. 1. A frame from the SeeThroughSmoke stereo benchmark.

arXiv 2026 · Preprint

Seeing Through Smoke: Surgical Desmoking for Improved Visual Perception

Real-time digital desmoking for robot-assisted surgery: a transformer with a physics-inspired desmoking head, a synthetic paired-data pipeline, and a real stereo smoke benchmark.

↳ SeeThroughSmoke — 5,817 paired stereo smoke images

PaperProjectDataset
SurgiSR 4K: A High-Resolution Endoscopic Video Dataset for Robotic-Assisted Minimally Invasive Procedures
Fig. 2. 1080p vs. native-4K detail, Figure 1 of the paper.

MELBA 2025 · MICCAI 2025 Oral

SurgiSR 4K: A High-Resolution Endoscopic Video Dataset for Robotic-Assisted Minimally Invasive Procedures

A public native-4K surgical endoscopy dataset supporting super-resolution, smoke removal, instrument detection, depth estimation, segmentation, novel-view synthesis, and VLM research.

↳ SurgiSR 4K — native-4K images and clips on Synapse

PaperarXivDataset
Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA
Fig. 3. Retrieval, reranking, and answer generation over PubMed.

CLEF 2025 · BioASQ Task 13b

Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA

A biomedical QA system combining dense retrieval, fine-tuned cross-encoders, GPT rerankers, and instruction-tuned LLM prompting over PubMed evidence.

PaperProceedings
Full publication list

§ 03

Experience

  1. Senior Machine Learning Engineer

    2026 — present

    Uber AV Labs

    • — Machine learning and perception for autonomous vehicles

  2. Senior Machine Learning Engineer

    2023 — 2026

    Intuitive Surgical

    • — GPU-accelerated video pipeline on NVIDIA Orin AGX: ultra-low end-to-end latency for 4K stereo surgical endoscopy with real-time depth estimation
    • — CUDA kernel development and TensorRT optimization — 1.8× throughput improvement
    • — Offline, privacy-centric, voice-enabled VLLM RAG system running in real time on edge devices
    • — Trained video world foundation models for surgical simulation with multimodal controls
    • — Led the iMED multi-endoscope dataset for sub-pixel reprojection accuracy
    • — Real-time auto-focus from eye tracking with Holoscan SDK, at 120 fps

  3. Machine Learning Engineer

    2022 — 2023

    Intuitive Surgical

    • — 4K stereo surgical endoscope prototype on NVIDIA Jetson Orin, integrated with Da Vinci 5 & Xi
    • — Led SurgiSR 4K dataset development for super-resolution research
    • — Next-gen camera sensor bring-up and ISP with downstream pattern recognition
    • — Image-stacking pipeline for surgical applications; Yocto OS and BSP development

  4. Machine Learning Intern

    2021

    Intuitive Surgical

    • — BiLSTM and AutoEncoder NLP models for automated error-log classification
    • — Reduced RMA with predictive failure modeling

  5. Software Engineer

    2017 — 2020

    Fresenius, NA R&D

    • — Class-2 life-critical medical software for hemodialyzers used by millions of patients
    • — Cloud telehealth infrastructure with Airflow, dbt, Databricks, and Snowflake; QNX RTOS development

§ 04

Toolbox

ML Training & Optimization
Transformers · Diffusion Models · LLM Fine-tuning (LoRA, QLoRA) · Mixed-Precision Training · Knowledge Distillation · Quantization (INT8/FP16) · TensorRT-LLM · vLLM
ML Frameworks & Infra
PyTorch (Multi-GPU) · CUDA · Triton · Distributed Training (DDP, FSDP) · MLflow · Weights & Biases · Model Serving at Scale
Multimodal & Vision
Vision-Language Models (VLMs) · CLIP · Cross-Modal Alignment · Video Understanding · RAG · LangChain · FAISS · Qdrant
Software & Platform
Snowflake · Apache Airflow · Kafka · dbt · Databricks · SQL · ETL/ELT Pipelines
Software & Cloud
Python · C/C++ · Docker · Kubernetes · AWS (SageMaker, S3, Lambda) · Azure · CI/CD (Jenkins, GitHub Actions)

§ 05 · Correspondence

Let's talk

Collaboration, research questions, or anything robotics- or ML-shaped — my inbox is open.

art_jiang@outlook.comLinkedIn