fig. 00 — α = 0.00
I build data and interaction tools that help people understand, evaluate, and work alongside AI.
1,017 kinds of work people brought to Claude in Feb 2026. Slide α from what the work is to how people split it with AI: handing it over (warm) or working alongside it (cool). Dot size = conversations. Laid out with my Class‑Constrained t‑SNE.
Data: Anthropic Economic Index (2026-03 release, MIT) · topics clustered from task text · occupations: O*NET via the 2025-03 release · hover or tap a dot
I am a software engineer at ByteDance, building developer tooling for Volcano Engine Ark, ByteDance’s Model-as-a-Service (MaaS) platform, and an AI-assisted video creation platform where creators and AI produce videos together.
Before that, I spent five years at Eindhoven University of Technology (PhD and postdoc) building algorithms and interactive tools that help people understand, compare, and debug AI models — and studying how they actually use them, from ML practitioners to clinicians.
I own the model fine-tuning, pricing, and usage modules of ArkCLI — the command-line tool for Volcano Engine Ark, ByteDance’s Model-as-a-Service (MaaS) platform. These are ArkCLI’s most-called modules. I also built ArkCLI’s skill library, which lets AI agents call ArkCLI directly.
I owned end-to-end development of the 3D module (Three.js), where creators set up virtual cameras in 3D scenes and produce videos together with AI. I also built the platform’s script editing module.
I work where AI meets data and people: tools that let people inspect and evaluate the data and models behind AI, and interfaces that let them question, steer, and work with it. Interactive visualization and user studies are my main methods.
Tools for finding what is wrong in the data and models behind AI: hard instances, abnormal clients in federated learning, long-running training experiments, and side-by-side model comparison.
PhD & postdoc, TU/e · Oxford e-Research Centre · Zhejiang UniversityInteractive systems and user studies: how clinicians and ML practitioners use AI tools, what-if questions with counterfactuals, LLM-written explanations of data, and a 3D interface where creators and AI make videos together.
EU SmartCHANGE · ELLIIT, Linköping · ByteDanceA CLI and skill library that let developers and AI agents use a Model-as-a-Service platform, and earlier, dashboards, a visualization component library, and open-source visualization for SynBioHub.
ByteDance · Cisco · Zhejiang University · Google Summer of Code
Visual tools to find and characterize the hard instances in a dataset — the samples deep neural networks struggle with — pointing to targeted model improvements.
A glyph-based visual analytics system for monitoring model pruning experiments over time, surfacing multi-metric trends across runs. Developed during a research visit to the Oxford e-Research Centre.
Interactive tools for probing AI model behavior, built in the EU project SmartCHANGE together with clinicians and ML researchers.
Integrating LLMs into visual analytics so that people get natural-language descriptions of the data clusters they select. A collaboration that began at the ELLIIT Focus Period in Sweden.
Under review · Preprint on SSRN
IEEE VIS Short Papers
VIS 2025 VISxGenAI Workshop
IEEE TVCG 30(1) · VIS 2023
Computer Graphics Forum 41(3) · EuroVis 2022
ACS Synthetic Biology 10(8)
Journal of Visualization 24(1)
IEEE TVCG · VAST 2018