AI agents / compute infrastructure / visual systems
Xinyu (Cindy) Zhang
Building AI-native systems for the physical world of compute.
Data scientist and software engineer working across AI agents, operational intelligence, and data center workflows. I like turning messy technical systems into tools, maps, and stories people can actually use.
research visuals
Readable maps for hard systems.
New public work will live here: visual essays, product comparisons, and interactive explainers about AI agents and data center operations.
visual research
Data Center Field Atlas
A Pudding-style explainer series on compute infrastructure, factory testing, cluster operations, and the strange metrics that decide whether machines are ready for production.
coming soonagent systems
Agent Evaluation Notes
Notes on tool use, traces, guardrails, handoffs, and the practical difference between impressive demos and dependable operational agents.
in progressindustry maps
Compute Ops Comparisons
Visual comparisons for vendors, workflows, failure modes, and reliability patterns across the data center ecosystem.
plannedselected work
Older experiments, kept as working traces.
Research projects, interactive tools, and LLM experiments from earlier work. New AI infrastructure case studies will join this archive as public material becomes available.
-
01
Glucose Interactive Browser
visualization
An interactive web-based tool to visualize and analyze glucose data
Open case study -
02
Spatial Statistic
AI4Science
Clustering Spatial Data after Hurricane Ian
Open case study -
03
Metric Voting
Math
Comparison of Different Randomnized Voting Algorithms
Open case study -
04
Gender Bias in LLMs
LLMs
Exploring the Gender Bias in LLMs (ChatGPTs, Gemini)
Open case study -
05
Auto Subtitle Generation
LLMs
Generating Subtitles for Videos using fine-tuned models
Open case study -
06
Solving Math Problems with LLM
LLMs
Solving phd level RA problems using fine-tuned ChatGPT
Open case study
how I build
Agents need taste, traces, and operating context.
Make invisible systems observable
Logs, tests, and telemetry become useful only when humans can see what changed and why it matters.
Build agents with traces
Good agents should leave behind plans, tool calls, uncertainty, and enough evidence to audit their decisions.
Explain before optimizing
The fastest way to improve a workflow is often to draw the system clearly enough that everyone can point at the bottleneck.
field notes
Writing from the build floor.
Notes on agents, automation, data tools, and the parts of engineering that get more interesting when they become visible.
- 2026-02-10» Claude Code 27 Day Retrospective
- 2026-01-26» Clawdbot Setup
- 2025-12-18» Adk Experience
- 2025-10-05» Ai Arrogance
- 2025-08-08» Spatial Statistic
- 2025-06-16» Glucose