ML systems researcher and engineer working on verifiable and efficient LLM infrastructure — from CUDA/Triton sparse-attention kernels and GPU cryptography up to agent security, routing, and edge feasibility.

First author on six peer-reviewed 2026 main-track conference papers (ICML, ECML PKDD, CIKM, ICICS, NLPCC ×2) plus nine workshop papers at ICML, ICLR, AAMAS, ProbML and C3NLP, with production backend and DevOps experience across AWS, Docker, and Kubernetes. Currently a Research Assistant at the CMU Infini-AI Lab.

CMU Infini-AI Lab — Research Assistant

PI: Prof. Beidi Chen · Pittsburgh, PA (Remote) · Mar 2026 – Present

• Core contributor to Vortex, a sparse-attention LLM inference framework on SGLang/FlashInfer; designed custom sparse-attention CUDA kernels with CUDA Graph capture for 2.7× end-to-end speedup.
• Built page-level cross-layer sparse-index sharing and a masked top-k kernel to cut redundant attention memory traffic across 1B–70B models.
• Built large-scale reproducible evaluation and benchmarking harnesses (multi-model sweeps, 30K+ runs, Docker/AWS, W&B tracking) — the backbone of the ICML 2026 red-team study.

Meetfood — Software Engineer Intern

Los Angeles, CA · May 2025 – Aug 2025

• Designed and shipped RESTful APIs with Node.js/Express/MongoDB — JWT auth, role-based access control, AWS Cognito, media uploads, and cross-device sync endpoints.
• Cut p95 latency of hot search/feed endpoints by ~40% (≈500 ms → 300 ms) via response caching and MongoDB query and index optimization.
• Built the video pipeline on AWS: S3 uploads, MediaConvert transcoding, CloudFront CDN delivery — cutting average video start time by ~50%; deployed on EC2 with auto-scaling and RDS under least-privilege IAM.
• Set up GitHub Actions CI/CD, cutting release turnaround from ~1 hour of manual steps to <10 minutes.

Tencent — Java Backend Engineer Intern

Shenzhen, China · Jul 2019 – Aug 2019

• Built backend modules for an internal messaging platform in Spring Boot with MyBatis; Redis caching of hot session data cut hot-path DB reads by ~60%.
• Refactored synchronous service calls into async CompletableFuture pipelines, reducing average response time by ~30% (≈200 ms → 140 ms).
• Deployed the team's microservices to Kubernetes (TKE) via declarative manifests with liveness/readiness probes and Prometheus/Grafana monitoring, cutting manual deploy time from ~30 min to <5 min.

Interruptr — Multi-Agent Vulnerability Scanner

Python, LangGraph, AutoGen · 2025–2026

• Multi-agent static-analysis pipeline deployable as a CI/CD security gate, scanning C/C++ for memory-safety bugs, race conditions and other CWE-classified vulnerabilities: three cloud experts plus one free local adversarial verifier.
• 77.2% F1 at $0.002/sample on NIST Juliet with 100% recall; the local verifier alone adds +10.3 pp precision (p<10⁻⁶, McNemar) and parallel execution gives a 3.0× speedup.

Timeline-Feed — Hardened High-Traffic Backend

TypeScript/Node.js, Kafka, Docker · 2025

• Production-grade feed backend with OWASP Top-10 hardening (XSS prevention, HMAC-signed cursors) and graceful-degradation circuit breakers.
• Sub-15 ms P95 at 99.99% availability; 30+ Prometheus metrics, Grafana, OpenTelemetry, and 80% test coverage across a 9-service Docker deployment.

geoffreywtech.me — Self-Hosted Research Portfolio

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

• Designed, deployed and operate this bilingual Wagtail CMS behind Nginx TLS termination on EC2, with content shipped as idempotent Django management commands rather than hand edits.
• Ran a full DevSecOps audit of the live stack (bandit/semgrep/pip-audit/trivy plus runtime probing): fixed a production DEBUG=True exposure and a CWE-601 open redirect, patched 132 CVEs over two audit rounds, and reached zero Django check --deploy warnings.

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

• Six first-author main-track acceptances in 2026 — ICML, ECML PKDD, CIKM, ICICS and NLPCC ×2 — spanning agent security, decision-theoretic serving, recommendation evaluation, verifiable inference, and citation integrity.
• Published open-source Rust crates hanfei-shu and hanfei-fa for GPU-accelerated zero-knowledge proving.
• Built and operate a self-funded edge–fog–cloud GPU testbed used for every 2026 paper: 2× Jetson Orin Nano (8 GB) + RTX 3080 Ti (16 GB) + workstation GPU server.

Open to relocation; availability immediate / negotiable. Reach me at zhaohui.geoffrey.wang@gmail.com — a PDF resume is linked at the top of this page.