About
Data Visualization Lead
Reliable software engineering in the AI era
I lead data visualization in the AI department at Hithink RoyalFlush, with a background in designing and delivering financial software. My work now centres on AI for finance and visualization: building interactive AI applications and exploring agent orchestration, evaluation, and the engineering infrastructure behind them. I focus on making model capabilities part of software we can verify and maintain.
I designed the core architecture of HiVis and worked with my team on Standard Chart and financial narrative visualization for company products. More recently, I have led the architecture and protocol design of Viz Platform, exploring multi-agent development, independent review, and recovery from interruptions in that project. Our HiVis research was published in the Journal of Visualization; work I co-authored has received a PacificVis team award and poster acceptances.
01
Turning financial data definitions, chart semantics, and interaction rules into protocols and components agents can use. Models produce structured chart specifications; code validates structure, semantics, and rendering. In Viz Platform, I am bringing business systems and agents onto a shared execution path.
Designing task contracts, roles, and execution state in AI Workflow to connect implementation, independent review, and integration checks. Isolated workspaces, task leases, and checkpoints handle concurrency and interruptions. The workflow is currently used for Viz Platform development on one machine.
Using AgentInfra to maintain reusable rules, skills, and workflows across coding clients, and Vis Agent Bench to record quality, time, token use, and human intervention under consistent task conditions. Failure records inform workflow improvements; evaluation samples are still being collected.
02
Paper
Journal of Visualization · 2026 · Published
Co-author · Core system architecture. A web-based financial visualization library supporting reusable charts, interactions, and cross-device adaptation.
Related conference records: ChinaVis 2025 paper and PacificVis 2025 poster, both concerning the HiVis system.
Poster
PacificVis · 2026 · Poster accepted
Co-author · Exploring automatic pictorial chart generation through context-aware asset selection and style control.
Contest
PacificVis 2025 Visual Data Storytelling Contest · Winner
Team award · Co-author. A data story examining the gender gap in economic engagement.
Patent
Granted · Co-inventor · May 2026 · ZL 2025 1 1138659.8 / CN 121071134 B
Co-inventor · Generating data visualization narratives based on narration.
03
I lead the architecture and protocol design, bringing AI chart rendering and Standard Chart into component, VizSpec, and runtime layers for developers, business systems, and upstream agents. Components can be used independently; acceptance of the combined release is still in progress.
I design task contracts and execution state for multi-agent development, using isolated workspaces, task leases, reviews tied to fixed commits, and a single integrator. Model fallback and checkpoints handle quota limits and interruptions. Used for Viz Platform on one machine; not formally released.
Explaining financial content through visuals and narration, including script planning, timing, and retries when generation fails.
I designed the core HiVis architecture and maintain chart components, interactions, and configuration interfaces shared across financial products. The related research was published in the Journal of Visualization.
I maintain rules, skills, workflows, and client adapters in one source repository. Manifests and generators keep client configurations consistent; installation previews, change records, and validation scripts support inspection and rollback.
My team and I designed the pipeline from analysis to interactive charts. VisualizationSpec constrains model output, with checks for structure, semantics, and rendering plus bounded repair attempts. Financial visualization knowledge is captured in a knowledge base and regression cases.
My team and I designed the evaluation method and platform to compare agents with consistent code, context, and tool permissions. Runs retain traces, automated checks, browser captures, and materials for anonymous review. The first three tasks validate the evaluation process as we collect more samples.
04
Hithink RoyalFlush, AI Department · Dec 2018 – Present
Hithink RoyalFlush, B2C Department · May 2016 – Dec 2018
Hithink RoyalFlush, B2C Department · Feb 2011 – May 2016
05
Spring 2027 · Enrollment pending
College of Computing, Georgia Institute of Technology
Sep 2007 – Jun 2011
Hangzhou Dianzi University, China
06
Three dedicated stories about TOEFL preparation, the system architect exam, and applying to OMSCS.
07
Outside work, I enjoy 3D technology, gaming, reading, and travel.
Write to discuss AI engineering, financial data applications, or visualization. Invitations to speak, join a podcast, collaborate on writing, or republish an article are welcome too.
Send an email