关于我

参与 toB 数据大屏类软件的交付与开发，面向政府及行业监测场景（如森林防火等可视化监管服务）。

• 参与跨校团队，研究并原型化基于 Unity 的「Echo（影子）」交互机制游戏；负责核心系统架构，降低交互延迟并稳定 WebGL 渲染。
• 在视觉表现与性能约束间权衡，于项目展示中汇报技术成果；课程成绩 A+（101/100），组内 第 1。

• 在 ACH 框架下完成生成式 AI 伦理专题研究报告，结合艺术与文化语境分析 LLM 应用边界、算法偏见与负责任部署，最终成绩 A+。

• 面向南科大学生的多服务 AI 助手：自然语言完成校园问答、多源日程整合、FileHub 文件管理与学习辅助；前端 React + Vite，SSE 流式对话，配套 Thought Trace 与 HITL 审批面板。
• 微服务后端：Go API Gateway / Orchestrator（gRPC）调度 Rust Agent Core 工具执行，Python LLM Service 统一对接 LLM；数据层 Postgres、Redis 会话摘要与 Qdrant 向量检索。
• 核心能力含校园百科 RAG、校历/Blackboard/个人 TODO 时间线、可编辑 Markdown Skills、周期任务 SSE 推送；Docker Compose 全栈部署。

• 复现并扩展 AHaBench 对齐流程：在 Qwen3-4B/8B/14B 与 Mistral-7B 上对比 baseline、few-shot、SFT、DPO 与 sequential SFT→DPO 五种 LoRA 训练方案。
• 采用 GPT-4o + DeepSeek-V4-Pro 双评委 0–6 量表评估；SFT→DPO 将 Mistral-7B 情感幻觉率降至 0.00，四模型平均分均 >5.1。
• 在 MMLU、GSM8K-CoT、ARC-Challenge 上验证对齐与通用能力基本正交，未见灾难性遗忘。

• 基于 ACS 加州/德州就业微观数据，针对 benchmark 中男女 FNR 差距（0.109 vs 0.198）设计三阶段缓解：子群逆频率重加权 → 五折 HistGBM 集成 → 分性别 Isotonic 校准 + 双阈值搜索。
• 在 17 项 FairnessIndicators 风格指标下 14/17 优于基准，达 Performance Level 5；整体 AUC 0.9017→0.9057，男女 FNR 分别下降 34% / 28%。
• 分析 equalized-opportunity 与 FPR–FNR 权衡，讨论分性别阈值在部署与合规场景下的适用边界。

• 在 SUSTechOS（xv6-riscv）上自研 Ext2/3/4，经 VFS 统一挂载，对接 virtio-blk 块设备，打通内核态到用户态的完整文件访问路径。
• 实现 inode / 目录项 / 块分配与缓冲缓存；扩展 Ext3 journal 保障元数据一致性，引入 extent 布局优化大文件寻址，支持索引块、稀疏文件等完整语义。
• 通过课程全量功能与压力测试，文件系统竞速 benchmark 排名 5/43（排名前11%）。

• 设计并实现 P2P 文件传输系统：文件按 chunk 分片、Peer 间对等传输，配合哈希校验保证完整性。
• 实现可靠数据传输机制，包括滑动窗口流量控制、累积 ACK、超时重传与快速重传（重复 ACK 触发），支持多路并发下载调度。
• 在优化版本中针对收发热路径降低协议栈开销，提升大文件、高并发场景下的传输吞吐与稳定性。

• 使用 Java 从零构建关系型数据库引擎，采用逻辑算子与物理算子分离架构，支持 INSERT 等全套基础 SQL 命令执行。
• 实现表存储、查询计划与算子流水线；逻辑层负责语义，物理层负责具体执行，便于扩展与协作开发。
• 实现 B+ 树索引，支持高效点查与范围检索。

设计并实现 RISC-V 单周期哈佛架构 CPU（64KB I/D、MMIO），于 EGO1 FPGA 部署运行；自研 Java 汇编器，将 .asm 编译为 ins.coe / data.coe 适配哈佛架构工具链。

用 Verilog 实现油烟机完整状态机（手势开关机、多档风速、照明、智能提醒、自清洁与参数配置），在 Artix-7 EGO1 开发板上验证。

使用 Java 从零实现三消类 Match-3 消消乐游戏。

• 将 cross-persona rewriting 的人格冲突建模为 Space-H 二维空间中的 structural stress，用冲突标量与方向向量刻画风格迁移。
• 搭建自动指标、LLM-as-Judge、CrEval pairwise 多通道评估，并以 mixed-effects regression 分析冲突与改写质量的关系。
• 研究 persona 表达与跨人格混合改写，对比 Prompt Engineering 与 激活向量编辑 两类干预手段，在离散/连续目标与机制实验臂下系统评估。

• 在 10 万 arXiv 元数据与 2 万 AI 标注全文上构建推荐系统，利用同质引用图作为对比学习信号。
• 使用 LoRA 微调 SciBERT，在留出测试集上推荐指标优于基线 4–5%。
• 叠加 RAG 检索层，支持自然语言论文查询与语义 + 图结构混合排序。

Software delivery and development for toB data-dashboard products—government and industry monitoring scenarios such as forest-fire surveillance and similar ops dashboards.

• Prototyped a Unity game around an “Echo (Shadow)” mechanic; owned core architecture for lower latency and stable WebGL rendering.
• Final showcase; awarded A+ (101/100), ranked 1st in cohort.

• Authored a research report on generative AI ethics within the ACH framework, examining LLM applications, cultural context, algorithmic bias, and responsible deployment — final grade A+.

• Multi-service campus AI assistant: natural-language Q&A, unified schedules, FileHub file workspace, and study workflows; React + Vite UI with SSE streaming, Thought Trace, and HITL approval panel.
• Microservices: Go API Gateway / Orchestrator (gRPC) orchestrates Rust Agent Core tool execution; Python LLM Service fronts LLM providers; Postgres, Redis session summaries, and Qdrant vector search.
• Campus encyclopedia RAG, multi-source calendar timeline, editable Markdown Skills, periodic tasks via SSE; full-stack Docker Compose deployment.

• Reproduced and extended the AHaBench alignment pipeline on Qwen3-4B/8B/14B and Mistral-7B under five LoRA regimes: baseline, few-shot, SFT, DPO, and sequential SFT→DPO.
• Dual-judge evaluation (GPT-4o + DeepSeek-V4-Pro, 0–6 rubric); SFT→DPO drives Mistral-7B affective-hallucination rate to 0.00 with avg. scores >5.1 on all four models.
• MMLU, GSM8K-CoT, and ARC-Challenge probes show alignment largely orthogonal to general capability—no catastrophic forgetting.

• On ACS CA/TX employment microdata, mitigated benchmark male/female FNR gap (0.109 vs 0.198) via subgroup reweighting, five-fold HistGBM ensemble, per-sex isotonic calibration, and dual-threshold search.
• 14/17 FairnessIndicators-style metrics beat the Keras benchmark (Performance Level 5); overall AUC 0.9017→0.9057, FNR down 34% / 28% for men and women.
• Analysed equalized-opportunity trade-offs between FPR and FNR, and deployment constraints for sex-specific thresholds.

• Built Ext2/3/4 on SUSTechOS (xv6-riscv): VFS mount layer, virtio-blk backing store, full kernel-to-userspace file I/O path.
• Inode & dentry management, block allocation, buffer cache; Ext3 journaling for metadata consistency; extent-based large-file mapping; indexed blocks and sparse files.
• Passed full course correctness and stress suites; ranked 5/43(top 11%) on the filesystem performance benchmark.

• Designed a P2P file-transfer system with chunk-based splitting, peer-to-peer delivery, and hash verification for integrity.
• Implemented reliable transfer: sliding-window flow control, cumulative ACKs, timeout retransmission, and fast retransmit on duplicate ACKs, with concurrent download scheduling.
• Optimized the send/receive hot path in an enhanced peer build to improve throughput and stability under large files and heavy concurrency.

• Built a relational DB engine from scratch in Java with separated logical and physical operators, supporting full basic SQL including INSERT.
• Logical operators handle semantics; physical operators execute plans—modular pipeline for storage, planning, and execution.
• Implemented B+ tree indexing for point and range lookups.

Designed a single-cycle Harvard RISC-V CPU (64KB I/D, MMIO) on EGO1 FPGA; built a Java assembler emitting ins.coe / data.coe for the Harvard toolchain.

Verilog state machine for a range hood—gesture power, fan speeds, lighting, smart reminders, self-clean, and settings—verified on Artix-7 EGO1.

Built a Match-3 tile-matching game from scratch in Java.

• Modeled persona conflict in cross-persona rewriting as structural stress in 2D Space-H, using conflict scalars and directional vectors for style transfer.
• Built multi-channel evaluation (automatic metrics, LLM-as-Judge, CrEval pairwise) and mixed-effects regression over conflict and rewrite quality.
• Studied persona expression and cross-persona hybrid rewriting; compared prompt engineering vs. activation vector editing under discrete, continuous, and mechanism experiment arms.

• Recommendation over 100K arXiv metadata and 20K annotated full texts using citation graphs as contrastive signal.
• LoRA-fine-tuned SciBERT beats baseline by 4–5% on held-out metrics.
• RAG layer for natural-language paper queries with hybrid semantic + graph ranking.