胡天乐HU, Tianle
求职方向:算法工程师 / 大模型应用(Agent)工程师 / 搜索与排序算法
Objective: Algorithm Engineer / LLM-Agent Application Engineer / Search & Ranking
电话Tel: 18118167910 ·
Email: hello@hutianle.com ·
GitHub: Horiz21 ·
博客Blog: blog.hutianle.com ·
主页Home
2027 届硕士(2027.04 毕业)· 2002.02 生 · 中共党员 · 英语 CET-6: 561
M.Sc., Class of 2027 (Apr) · b. 2002.02 · CPC member · English: CET-6 561
2027 届计算机学术型硕士,三段实习:华为(AI Agent 研发)、百度(搜索算法,成果上线百度首页)、中科院软件所(强化学习方向)。GitHub 400+ Star 开源项目核心贡献者,发明专利主要撰写人(第三发明人,已公开)。聚焦多模态信息抽取、VLM/LLM 应用、搜索重排序与 Agent,具备从数据构建、模型训练到上线评估的完整链路经验。
M.Sc. candidate (graduating Apr 2027) with three internships: Huawei (AI-agent R&D), Baidu (search algorithm — shipped to the Baidu homepage) and Institute of Software, CAS (reinforcement learning). Core contributor to a 400+ star open-source project; primary author of a published invention patent (3rd inventor). Focused on multimodal information extraction, VLM/LLM applications, search re-ranking and agents — full pipeline experience from data construction and training to online evaluation.
南京理工大学 · 计算机科学与技术 · 学术型硕士(在读)Nanjing University of Science & Technology — M.Sc., Computer Science
2024.09 – 2027.04
90.4/100,专业排名 23/335(前 10%)· 硕士一等奖学金 ×290.4/100, rank 23/335 (top 10%) · First-class scholarship ×2
南京信息工程大学 · 计算机科学与技术 · 工学学士 / 荣誉学士Nanjing University of Information Science & Technology — B.Eng. (Honours), CS
2020.09 – 2024.06
GPA 4.2/5.0,专业排名 2/221(前 1%)· 国家奖学金 · 校长奖学金 · 优秀毕业生GPA 4.2/5.0, rank 2/221 (top 1%) · National Scholarship · Presidential Scholarship · Outstanding Graduate
华为 · 云计算业务单元 · 智能体研发工程师(实习)Huawei — Cloud BU, AI Agent R&D Engineer (Intern)
2026.06 – 2026.09
本体驱动的自然语言数据分析 Agent · 华为南京研究所Ontology-driven natural-language data-analysis agent · Nanjing Research Institute
- 负责 Agent 技术预研与可运行 Demo:建模 60+ 本体类、映射 6 张业务表,Ontop 经 OBDA/R2RML 将 SPARQL 改写为 SQL,SKOS 对齐业务词表,打通"自然语言 → 语义层 → 数据层"查询链路。Owned pre-research and a working demo: modeled 60+ ontology classes mapped to 6 business tables; SPARQL rewritten to SQL via Ontop (OBDA/R2RML); vocabulary aligned with SKOS — a full NL → semantic → data query chain.
- 提出辅助类剪枝与口径元数据机制,解决多粒度/多表命中导致的同一业务数据重复统计;设计查询前边界校验链路,拦截语义不连通与超能力范围请求并返回原因。Proposed auxiliary-class pruning with caliber/granularity metadata, fixing duplicate counting from multi-table hits; designed pre-query boundary validation that intercepts disconnected or out-of-scope requests with reasons.
- 将场景定义从固定 SKILL 迁移入本体,运行时动态发现(覆盖 3 场景、约 10 项能力),并开发 6 个通用 + 约 10 个场景私有工具(MCP / Agent Skill)。Moved scenario definitions from fixed skills into the ontology for runtime discovery (3 scenarios, ~10 capabilities); built 6 general plus ~10 scenario-specific tools (MCP / Agent Skill).
- 成果:含历史失败案例的 50 题对照,正确率 42/50 → 47/50(84% → 94%);交付可运行产物与完整预研材料,推进企业本体建设探索(对标 Palantir Ontology)。Result: on a 50-question suite incl. historical failures, accuracy rose from 42/50 to 47/50 (84% → 94%); delivered a runnable artifact and full pre-research report, advancing the enterprise-ontology initiative (cf. Palantir Ontology).
百度 · 搜索策略部 · 算法工程师(实习)Baidu — Search Strategy Dept., Algorithm Engineer (Intern)
2025.11 – 2026.02
生成式搜索视频多模态重排序(第一负责人)· 事实校验可信度评分Video multimodal re-ranking for generative search (first owner) · credibility scoring for fact-checking
- 主导"故事图"视频多模态重排序:基于时间对齐的 ASR/OCR 语义分段与时间密度调节,九宫格表征在固定输入预算下最大化保留叙事逻辑;对比成对/成组/独立打分选型,使评价成本不随候选规模增长。Led the "story-graph" video multimodal re-ranking: time-aligned ASR/OCR semantic segmentation with time-density adjustment; a 3×3-grid representation preserves narrative logic under a fixed input budget; selected among pairwise / groupwise / independent scoring so evaluation cost stays fixed.
- 以 200 对人工偏好基准选定主/副裁判模型,经加权投票从约 1 万条候选构建 约 6700 条高质量偏好微调数据;基模 int4 量化 Qwen2.5-VL-72B,内部平台 LoRA 微调。Selected primary/secondary judge models against a 200-pair human benchmark and built ~6,700 preference training samples via weighted voting from ~10k candidates; LoRA fine-tuning on int4-quantized Qwen2.5-VL-72B.
- 模型上线百度搜索首页:千分之一流量五天测试中,跳转点击率与视频播放留存绝对提升约 0.2 个百分点;相关算法写入发明专利。Shipped to the Baidu Search homepage: in a 5-day test on 0.1% traffic, click-through and video-play retention improved by ~0.2 percentage points absolute; the method was written into an invention patent.
- 担任专利 CN122470267A 主要撰写人(第三发明人)。Primary author (3rd inventor) of patent CN122470267A.
- 实习末期参与 ReAct 事实校验:实现单文档 / 单论断 / 召回充分性 / 总结覆盖冗余四层可信度评分,规则硬信号与模型语义信号混合,按查询意图动态调权;案例验证拦截虚假"科比复活"内容(两周 Demo)。Also built a four-layer credibility scoring prototype (document / claim / recall-sufficiency / summary-coverage) in the ReAct fact-checking flow — rule-based hard signals mixed with model semantics, weights adjusted by query intent; demo intercepted a fabricated "Kobe resurrection" claim (2-week demo).
中国科学院软件研究所 · 算法工程师(实习)Institute of Software, Chinese Academy of Sciences — Algorithm Engineer (Intern)
2025.05 – 2025.11
多智能体仿真 · 中控策略强化学习训练 · 训练调试系统(项目脱敏表述)Multi-agent simulation · RL training of the central-control policy · training/debug toolchain (sanitized)
- 负责中控策略训练:Qwen 系 7B 基座 + LoRA,verl 承载 GRPO、vLLM rollout 采样(8×4090),Python 侧经 gRPC 对接 Java 仿真后端,累计约 400 步更新、约 2.5 万局仿真;以 GRPO 短指令替代行为树,完成资源调度、冲突仲裁与全局策略切换。Owned central-control policy training: Qwen-7B-class base + LoRA, GRPO via verl with vLLM rollouts (8×4090), gRPC to a Java simulation backend — ~400 update steps over ~25k episodes; short GRPO commands replaced behavior trees for resource scheduling, conflict arbitration and global policy switching.
- 设计局末分按折扣回填决策步的混合奖励;定位并治理拖延超时、格式合规无实义、重复指令刷步数三类奖励黑客。Designed a hybrid reward backfilling end-game scores to decision steps with discounting; diagnosed and fixed three reward-hacking patterns (stalling, format-only compliance, repetition).
- 独立开发 Flask 训练调试前端:实时态势推送、整局回放、状态注入与时间轴对齐,被团队持续使用约三个月并用于项目演示。Independently built a Flask debugging frontend — real-time state push, full-game replay, state injection and timeline alignment — used by the team for ~3 months and in project demos.
- 本地部署 Qwen,搭建覆盖百余份内部文档的 RAG(512/64 token 切块、top-5 检索)与约 100 条人工问答评测,用于参数调优。Deployed local Qwen with RAG over 100+ internal documents (512/64-token chunks, top-5 retrieval) plus a ~100-question human evaluation for tuning.
- 成果:项目内测试相对行为树取得优势,低难度胜率八成以上。Result: outperformed behavior-tree baselines in internal tests, with an 80%+ win rate at low difficulty.
更多个人项目More personal projects
P5CCG(小红书 5000+ 点赞)、Lambda Launcher 等近 10 个个人开源 / 工具项目,详见 GitHub 与技术博客。P5CCG (5,000+ likes on Xiaohongshu), Lambda Launcher, and nearly 10 personal open-source / tool projects — see GitHub and my blog.
南京信息工程大学计算机学院科学技术协会 · 主席President, Science & Technology Association, School of Computer Science, NUIST
2022 – 2024
组织或协办学院与学校竞赛十余场(含校级算法竞赛团体赛)、前沿科技讲座十余次(主持 3 次);任期内招新 40 余人,参与部门选拔调整。Organized 10+ contests (incl. a university-level team algorithm contest) and 10+ tech lectures (hosted 3); recruited 40+ members; led member selection and adjustments.