Skip to content
View Zhuoxi2000's full-sized avatar

Block or report Zhuoxi2000

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Zhuoxi2000/README.md

I build stream systems moving 1e9+ events/day by day, and break AI agents for science by night — deterministic, annotation-free health checks(体检) for the agentic era. My rule for both jobs is the same: if a claim can't be checked by an oracle, it's a vibe, not a guarantee.

01 · BUILD × BREAK

project what it is
agent security showroom two identical e-commerce agents — one naked, one behind my oracle guardrail. you pick the injection; watch one leak and one refuse, deterministically
OracleBench annotation-free agent-security leaderboard — DeepSeek / Qwen / GLM / Kimi next to frontier models, one command to reproduce
ServeSense health checks for my tennis serve: in-browser pose estimation grading toss consistency & contact point. agents get checkups; so does my serve
wild hawk-eye offline pipeline (TrackNet + court homography) that turns phone video of my matches into trajectory JSON — all compute happens before serve time
ed-w.com parent process of this README. one hand-written HTML file; no framework survived the code review

● running in prod   ◌ in the lab

02 · MEASURE

Benchmarks I'm building, all on one principle: security and reliability claims should be decided by a deterministic oracle, not a rubric.

  • CanaryBench — first deterministic, annotation-free benchmark for indirect-prompt-injection resistance — accepted, to appear
  • AgentLeak-Bench — incidental privacy leakage of tool-using agents, through the lens of contextual integrity
  • StreamReason-Bench — can LLMs reason about event-time streaming semantics? (my day job, mirrored into a benchmark)
  • Lost in the Trajectory — sequence-position bias when LLM judges do step-level error attribution
  • plus a software-supply-chain × LLM series (SBOM / VEX / CVSS / SLSA)

03 · OFF-COURT

$ tail -f /var/log/edson/offcourt.log
[serve ] tossing balls at ServeSense until the toss stops wobbling
[courts] tennis · badminton — the original adversarial evaluation
[winter] skiing
[iron  ] barbell, compound lifts only

this page is static — the demos are not: go inject something
banner compiled by tools/gen_hero.py · typeface: JetBrains Mono (OFL) · even this README has a build step

Pinned Loading

  1. tracelake tracelake Public

    Python

  2. iceberg-python iceberg-python Public

    Forked from apache/iceberg-python

    PyIceberg

    Python

  3. polaris polaris Public

    Forked from apache/polaris

    Apache Polaris, the interoperable, open source catalog for Apache Iceberg

    Java