Skip to content
View Quarkgluonmixture's full-sized avatar
5 followers · 1 following

Achievements

Achievement: Pull Sharkx3Achievement: Pair ExtraordinaireAchievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pull Sharkx3Achievement: Pair ExtraordinaireAchievement: QuickdrawAchievement: YOLO

Highlights

  • Pro

Block or report Quarkgluonmixture

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.

Content in all repositories owned by your account will be closed.
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
Quarkgluonmixture/README.md

Hi, I'm Jiaming Wei

Research engineer working on AI evaluation, red teaming and agent reliability.

I study how to evaluate agentic systems when the evidence itself is unreliable: task outcomes vary across reruns, automatic judges make systematic errors, and a final success label can hide failures in trajectories, tools, providers or the evaluation infrastructure.

  • 🎓 MSc Artificial Intelligence for Sustainable Development, UCL Computer Science (2025–26) · BEng Automation, Xi'an Jiaotong University
  • 🧪 Research Intern, Holistic AI, London (Jun–Sep 2026): LLM / agent red-teaming and evaluation
  • 📄 Routing Is Least Learnable Where It Is Most Valuable: Bounds on Representation Routing for Web Agents — accepted at the EMNLP 2026 Workshop REALM (OpenReview)

Selected work

Project What it is
Cost-Aware-Routing-for-Web-Usage-Agents Preregistered study of which page representation (DOM / Set-of-Marks / vision) a web agent should read, across 3 model families × 6 observation modes. Main result: representations carry real, complementary value, but that value is hardest to predict exactly where it matters most.
redteam-under-test Local-first LLM red-team measurement stack: target-conditioned attacks, 103 end-to-end-validated plugins, and the judge itself measured against independent gold labels.
agent-redteam-lab Tool-using agent security: a read-only MCP trace analyzer for boundary crossing, data egress, destructive actions and confused-deputy behaviour, plus a replay harness.
coding-agent-guardrails Rules, hooks and skills that keep AI coding agents (Claude Code, Codex CLI) reliable: git guardrails, evidence before a verdict, session handover, cross-model review.
constrained-agent-runner Control plane for a web chat agent to run constrained, auditable jobs on your own machine, with GitHub as durable transport and hard capability boundaries.
FinQA Controlled post-training attribution on Qwen3 (0.6B→14B): separating protocol alignment, answer-format SFT, explicit reasoning and GRPO before crediting a score change to capability.

Links · Portfolio · Research CV · Poster & showcase · LinkedIn · jiaming.wei.ai@outlook.com · 中文主页

Pinned Loading

  1. Cost-Aware-Routing-for-Web-Usage-Agents Cost-Aware-Routing-for-Web-Usage-Agents Public

    Python 2 1

  2. FinQA FinQA Public

    Thinking-aware baselines & low-data LoRA/QLoRA post-training on FinQA — a controlled Qwen3-4B vs Qwen3-8B numerical-reasoning study.

    Python 4 4

  3. agent-redteam-lab agent-redteam-lab Public

    Python 2

  4. redteam-under-test redteam-under-test Public

    Local-first LLM red-teaming: target-conditioned attacks, audited technique fidelity, and a judge scored against independent gold labels. Runs with zero egress.

    JavaScript 2

  5. coding-agent-guardrails coding-agent-guardrails Public

    Rules, hooks and skills that make AI coding agents (Claude Code, Codex CLI) reliable: git guardrails, evidence-before-verdict discipline, session handover, cross-model review.

    Shell 1

  6. constrained-agent-runner constrained-agent-runner Public

    Control plane for a web chat agent to run constrained, auditable jobs on your own machine — GitHub as durable transport, hard capability boundaries, recoverable runs.

    PowerShell 1