机器学习系统方向的研究员与工程师，专注于可验证、高效的大模型基础设施——从 CUDA/Triton 稀疏注意力内核与 GPU 密码学，到智能体安全、请求路由与边缘部署可行性。

2026 年以第一作者身份发表 6 篇主会论文（ICML、ECML PKDD、CIKM、ICICS、NLPCC ×2），另有 9 篇 ICML / ICLR / AAMAS / ProbML / C3NLP workshop 论文；同时具备 AWS、Docker、Kubernetes 上的生产级后端与运维经验。现任 CMU Infini-AI Lab 研究助理。

CMU Infini-AI Lab — 研究助理

导师：Beidi Chen 教授 · 美国匹兹堡（远程） · 2026 年 3 月 – 至今

• 稀疏注意力大模型推理框架 Vortex 的核心贡献者（基于 SGLang/FlashInfer）；设计定制稀疏注意力 CUDA 内核并配合 CUDA Graph 捕获，实现端到端 2.7× 加速。
• 实现页级跨层稀疏索引共享与 masked top-k 内核，在 1B–70B 模型上削减冗余注意力访存。
• 构建大规模可复现的评测与基准测试框架（多模型扫描、3 万余次运行、Docker/AWS、W&B 追踪），支撑 ICML 2026 红队研究。

Meetfood — 软件工程实习生

美国洛杉矶 · 2025 年 5 月 – 2025 年 8 月

• 使用 Node.js/Express/MongoDB 设计并交付 RESTful API——JWT 鉴权、基于角色的访问控制、AWS Cognito、媒体上传与跨设备同步接口。
• 通过响应缓存与 MongoDB 查询/索引优化，将核心搜索与信息流接口的 p95 延迟降低 约 40%（约 500 ms → 300 ms）。
• 在 AWS 上搭建视频流水线：S3 上传、MediaConvert 转码、CloudFront CDN 分发，平均起播时间降低 约 50%；服务部署于 EC2（自动扩缩）与 RDS，遵循最小权限 IAM。
• 搭建 GitHub Actions CI/CD，将发布周期从约 1 小时的人工操作压缩到 10 分钟以内。

腾讯 — Java 后端开发实习生

中国深圳 · 2019 年 7 月 – 2019 年 8 月

• 基于 Spring Boot 与 MyBatis 开发内部消息平台后端模块；Redis 热点会话缓存使热路径数据库读取减少 约 60%。
• 将同步服务调用重构为异步 CompletableFuture 流水线，平均响应时间降低 约 30%（约 200 ms → 140 ms）。
• 通过声明式清单（基础设施即代码）将团队微服务部署至 Kubernetes（TKE），配置存活/就绪探针与 Prometheus/Grafana 监控，人工部署耗时从约 30 分钟降至 5 分钟以内。

Interruptr — 多智能体漏洞扫描器

Python、LangGraph、AutoGen · 2025–2026

• 可作为 CI/CD 安全门禁部署的多智能体静态分析流水线，检测 C/C++ 内存安全缺陷、竞态条件等 CWE 分类漏洞：三个云端专家 + 一个零边际成本的本地对抗验证器。
• 在 NIST Juliet 上达到 77.2% F1、$0.002/样本、100% 召回；仅本地验证器一项即带来 +10.3 个百分点的精确率提升（McNemar 检验 p<10⁻⁶）。

Timeline-Feed — 高流量加固后端

TypeScript/Node.js、Kafka、Docker · 2025

• 生产级信息流后端，完成 OWASP Top-10 加固（XSS 防护、HMAC 签名游标）与优雅降级熔断。
• P95 低于 15 ms、可用性 99.99%；9 个服务的 Docker 部署中包含 30+ Prometheus 指标、Grafana、OpenTelemetry 与 80% 测试覆盖率。

geoffreywtech.me — 自建研究作品集网站

Django 5.2、Wagtail、Docker、Nginx、AWS EC2 · 2025–2026

• 独立设计、部署并运维本双语 Wagtail CMS（Nginx TLS 终止 + EC2），内容通过幂等的 Django 管理命令发布，而非手工编辑。
• 对线上环境执行完整 DevSecOps 审计（bandit/semgrep/pip-audit/trivy 结合运行时探测）：修复生产环境 DEBUG=True 暴露与 CWE-601 开放重定向，两轮审计共修补 132 个 CVE，最终 Django check --deploy 零告警、pip-audit 无已知漏洞。

First author on all. Full list with PDFs, figures and posters: geoffreywtech.me/publications

Main track

• Persistent Semantic Entities in Tool-Augmented LLM Systems — ICML 2026. Formalized a new class of LLM-agent vulnerability: implicit state in tool registries and event systems that persists across sessions and propagates across agents, invisible to standard prompt-injection monitoring. 34,000+ trials over 20 models from 9 vendor families (1.5B–1T); all susceptible (median 70% contamination), uncorrelated with scale (r²=0.06); name-binding dominant (Cohen's d=3.26); cross-model validation defense eliminates the vector (median 100% reduction).
• Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems — ECML PKDD 2026. Cast when to invoke an expensive cloud LLM as risk-based sequential stopping; unified framework subsuming event-triggered / optimal-stopping / SPRT / CUSUM / bandit triggers (6 theorems, incl. O(√(T log T)) regret). <5% miss at 6.2% invocation (~10× fewer calls); 63% lower total cost than CUSUM.
• NanoZK: Layerwise Zero-Knowledge Proofs for Verifiable LLM Inference — ICICS 2026. Made outsourced LLM inference cryptographically verifiable via layerwise ZK proofs (Halo2/IPA) with a custom GPU multi-scalar-multiplication CUDA kernel, released as the Rust crate hanfei-shu. 7.6× MSM speedup; 620 ms prove / 2.4 KB proof / 2 ms verify on LLaMA-3-1B at zero accuracy loss.
• The Recall Ceiling: Why LLM Recommendation Reranking Fails in Practice — CIKM 2026. Showed the oracle evaluation protocol used for LLM rerankers overstates NDCG@10 by 92–95%, and proved a deterministic ceiling: expected NDCG@k cannot exceed recall at the reranker's own candidate-window size under leave-one-out evaluation. Realistic retrieval covers only 2–19% of relevant items at K=100 across eight datasets in three domains.
• IntegriRef: A Five-Layer Bayesian Framework for Cross-Domain Reference Integrity Verification — NLPCC 2026. Five-layer pipeline over 62 registry adapters fusing 21 typed signals with Bayesian likelihood-ratio scoring: 99.9% recall on 1,215 hallucinated references and 98.3% on 5,488 retracted papers; L2 semantic verification reaches 93.5% precision against 89.1% human agreement.
• When Retrieval Systems Encode Wrong Entity-Context Bindings — NLPCC 2026. Index-level adversarial memorization: false entity–context associations need not be learned in pretraining, yet are preserved and amplified by the retrieval index. Multilingual case study over 1,535 articles; an evidence-grounded rewriting prototype removes 83.3% of pollution where the best filter reaches 20.8%.

Workshop

• When Individually Calibrated Models Become Collectively Miscalibrated — ProbML 2026 (PMLR archival). Collective miscalibration under strategic prediction aggregation for edge–cloud inference, with an incentive-compatible (VCG/Brier) mechanism. 0/6 exploits; false-negative rate 0.027 vs 0.045–0.067 for neural baselines; 4× more data-efficient; sub-2 ms at 200 agents.
• Interruptr — AAMAS 2026 SE Workshop · arXiv:2604.21282
• AgentTrace — AIWILD @ ICLR 2026 · arXiv:2603.14688
• Universe Routing — LLA @ ICLR 2026 · arXiv:2603.14799
• NanoZK (workshop version) — VerifAI @ ICLR 2026
• Block-Level Weight-Space Structure Persists Under Post-Training — Weight-Space Symmetries (WSS) @ ICML 2026. Post-training rewrites every tensor yet preserves block-level geometry (cosine > 0.99), enabling 18–48% GPU memory savings across co-resident model variants.
• Learning from World Feedback — RLxF @ ICML 2026. Dynamics-based uncertainty penalties increase collisions from 26% to 34%; replacing the model-internal proxy with world-feedback signals cuts them to 1–14% without retraining.
• From Risk Scoring to Risk Allocation — SRAS @ ICML 2026. Monitoring in multi-agent systems recast from per-state scoring to combinatorial allocation: a QUBO whose Pareto frontier contains greedy top-K, MMR and k-DPP as fixed operating points. +24% to +66% diversity over greedy from n=15 to n=200; a learned λ policy reaches 99.5% of an oracle grid search. OpenReview
• TabletCraft — C3NLP @ ACL 2026 · bidirectional Akkadian NMT and cuneiform rendering; 49.1 / 48.5 BLEU, the first published quantitative result in the reverse direction. pip install cuneiscribe

• 2026 年获得 6 篇第一作者主会录用——ICML、ECML PKDD、CIKM、ICICS 与 NLPCC ×2，覆盖智能体安全、决策论驱动的服务调度、推荐系统评测、可验证推理与引用完整性。
• 发布开源 Rust crate hanfei-shu 与 hanfei-fa，用于 GPU 加速的零知识证明。
• 自建并运维边缘–雾–云 GPU 测试平台，支撑全部 2026 年论文实验：2× Jetson Orin Nano（8 GB）+ RTX 3080 Ti（16 GB）+ 工作站 GPU 服务器。