Evaluation
Human-centered evaluation for AI agents
Benchmarks, rubric design, and product telemetry for measuring whether agentic systems behave well outside curated demos.
Zhe Sun
Name pronunciation: “Juh Soon”
I work at the boundary of academic AI research and product execution: turning models, evaluations, and human workflows into systems people can trust, inspect, and use.
Featured Research
Research threads I want the site to make easy to scan, cite, and discuss.
Interaction
Interfaces for controllable model behavior
Design patterns that expose uncertainty, provenance, and revision control without making advanced AI tools feel heavy.
Systems
Applied AI systems from prototype to deployment
Translating papers into resilient product architectures: retrieval, feedback loops, eval harnesses, and monitoring.
AI Product Work
Selected product directions for translating research capability into useful workflows.
Research tooling
AI research copilot for literature synthesis
A product concept for collecting papers, extracting claims, comparing evidence, and producing citable research notes.
Enterprise AI
Model evaluation dashboard for product teams
Dashboards, datasets, and release gates for teams deciding when an AI feature is good enough to ship.
Workflow design
Human-in-the-loop automation patterns
Review queues, editable drafts, confidence thresholds, and intervention models for practical AI operations.
Knowledge systems
Knowledge infrastructure for expert teams
Structuring internal knowledge so retrieval and generation systems can stay grounded, current, and auditable.
Collaboration Areas
LLM Evaluation
Agent UX
AI Safety
Applied ML
Product Strategy
“The work I want this site to foreground is research that survives contact with messy product reality: measurable, inspectable, and useful to the people making decisions.”
Focus
Reusable building blocks for the academic/product profile.
Research Agenda
Concise positioning for publications, preprints, talks, and ongoing questions.
View researchProduct Portfolio
Case-study slots for shipped tools, prototypes, evaluations, and strategy work.
View productsPublic Writing
Essays that connect model capability, evaluation culture, and product judgment.
Read writingAdvising
Clear entry points for collaboration with labs, founders, and product teams.
Start a conversationWriting
What AI product teams should measure first
A practical note on choosing eval metrics before the prototype starts shaping the research question.
The case for slower agent demos
Why legibility, interruption, and repair often matter more than an impressive autonomous run.
Designing interfaces around uncertainty
Patterns for making model confidence useful without burying users in instrumentation.
Notes on research taste in applied AI
How to keep academic rigor alive when the work also has to ship.