Me

About Me

hyacehila is my long-term online ID. It comes from hyacinth, my favorite plant. I later reshaped it into a lighter, more name-like form: hyacehila. The ending -hila / -ila gives it a small, airy, fictional texture, so to me it is not only a username, but also a hyacinth sprite living inside complex structures.

This ID is close to how I understand technology: moving through dense systems, toolchains, workflows, and uncertain environments to find a natural, explainable path that actually solves the problem. I care about general problem-solving patterns and about whether technology can transfer and generalize across real scenarios.

Today I mainly focus on AI Agent deployment and Evaluation, which I see as two of the most important technologies for bringing large models into engineering practice and industrial pipelines. I study both Single-Agent and Multi-Agent systems: how they are designed, evaluated, constrained, and eventually embedded in real business workflows rather than left at the benchmark level. Writing commercial fiction is a side interest.

You can call me Julian or Jules.

Intern

AI Agent R&D Engineer (Intern)

NetEase Interactive Entertainment (Shanghai) · May 2026 -- Present

  • Game UI Production-Chain Agent: Researching an agent tool around the planning GUI, UIP, and engineering development workflow to generate NeoX-side .uiprefab files from Figma/PSD designs. The system uses multimodal embedding retrieval, project template libraries, and development guidelines to inject component hierarchy, asset binding, naming conventions, and engine-validity constraints, enabling UI reconstruction and UIP integration within existing project conventions while continuing to tackle state combinations, animation consistency, and complex-interface coherence.
  • UI Program Automation & Real Feedback Loop: Developed Agent Skills for the NeoX engine and internal UI editor, connecting project knowledge, server-side MCP, and client-side MCP to turn one-shot code generation into a multi-turn generate-run-observe-fix loop. Through simulated clicks and real UI feedback, the agent can locate UI logic, asset binding, and interaction issues, allowing planners to complete initial UI programming without reading code while engineers focus on review and final merge control.
  • Project-Level Knowledge Base & Intelligent Q&A: Designed and implemented a self-evolving knowledge base based on LLM Wiki and Agentic RAG to address stale documentation and implicit knowledge loss caused by long-term iteration of the in-house engine and UI editor. The system converts scattered and orally transferred project experience into a searchable, traceable, and dynamic context layer, substantially lowering onboarding cost and technical communication overhead inside the project team.

Algorithm Researcher (Intern)

NSFOCUS Technology (Wuhan) · Dec 2025 -- Mar 2026

  • Vulnerability Mining Agent & CodeQL Verification Loop: Built a single-agent harness for code vulnerability discovery, organizing vulnerability-intelligence retrieval, source-code localization, taint-flow modeling, CodeQL query generation, and engine validation into a traceable multi-round analysis loop. The work emphasized state representation, tool-calling protocols, and verification feedback over preset multi-role decomposition, enabling the model to iterate around candidate source/sink pairs, failed paths, tool outputs, and validation results while reducing path dependence and false-positive conclusions in long-chain analysis. CodeQL taint-analysis results served as the external verifier that provided final feedback.
  • Training Data & Agent Trajectory Curation: Built a vulnerability-data cleaning, entity verification, and annotation pipeline from GitHub open-source projects and internal databases; extracted and verified 8,000+ CVE entities, cleaned about 4,000 high-quality samples, completed multi-dimensional CWE/OWASP labeling, and controlled the language and vulnerability-type mix. From agent executions, distilled 2,500 high-confidence tool-use SFT trajectories and 300 verified taint-flow records for cold start, evaluation-set construction, reward design, and follow-up RL training exploration.

Research

  • Unveiling the Drivers of PTSD: An Interpretable Machine Learning Approach with SHAP International Conference on Intelligent Computing and Data Analysis 2025 ; EI DOI

Awards & Certificates

  • First Prize in Shaanxi Province, National College Students Statistical Modeling Competition
  • Third Prize National, SAS China University Data Analysis Competition
  • Second Prize, Mathematical Contest in Modeling (MCM/ICM)
  • CET4: 510 | CET6: 513