Founder of Nodexi Agentics, LLC — an agentic-intelligence research company building agents that are general and safe at the same time. PhD student at Fudan University, visiting student at HKUST. I study how to break and defend LLM-based agents — tool-invocation security, response-path integrity, post-alignment tampering, and endogenous safety mechanisms. A second line of work asks whether our measurements of AI safety actually measure what we claim: scoring polarity, confidence intervals for ratio metrics, and budget-matched comparisons.
Everything I research, I also ship. The agent infrastructure below runs my own daily workflow — multi-model orchestration, capability-scoped tool execution, and sandboxed agent runtimes.
Research: LLM Security • Agent Safety • Evaluation Methodology • Kernel Fuzzing • AI-enabled Vulnerability Detection
| Project | Description |
|---|---|
| TIPExploit | Empirical risk assessment of tool-invocation prompts in LLM agentic systems (under review) |
| aap-audit | Artifact for "Measuring the Wrong Thing: Internal Harmfulness Scores Anti-Rank Successful Jailbreaks" |
| RTA | Response-path integrity attack & defense evaluation platform for BYOK LLM agents |
| dhr-security-platform | Attack & Dynamic Heterogeneous Redundancy (DHR) defense benchmark for autonomous-driving perception |
| syzkaller (fork) | Kernel fuzzer enhanced with LLM-assisted mutation for improved coverage |
| Project | Description | Stars |
|---|---|---|
| pm-mode-plugin | Claude Code PM-Mode — two-tier cross-model / cross-CLI dispatch with hook-based approval, dashboard, health checks, 34+ commands |  |
| agent-your-agent | Multi-agent orchestration framework — routes tasks to the best model (Claude/Deepseek/GPT) via file-system protocol |  |
| agent-forge | Workbench to forge and govern enterprise AI agents — capability-scoped operations, dataflow visualization, full action auditing (CaMeL dual-LLM pattern) |  |
| log-adhd | Action-first output shaping for Claude Code that holds across a whole session — four intensity levels + a shape-compliance meter |  |
| llm-roundtable | Multi-LLM structured debate platform — moderator / expert / critic agent roles |  |
| camel-plugin | ClawsGO client — skill synchronization against the CaMeL platform |  |
| Project | Description | Stars |
|---|---|---|
| SoulByte | Turn WeChat chat history into AI training datasets and personal knowledge bases — 72-hour context construction, contact-graph management, LLM-based quality scoring |  |
| mutilated_text_recognition | Attention-enhanced deep model for recognizing mutilated / damaged text (1 software copyright) |  |
| NexusAI-Hub | Unified multi-provider AI model management with OpenAI-compatible APIs, cost tracking and usage analytics |  |
| CaMeL-docs | Documentation site for the CaMeL ecosystem |  |
