Aojie (Justin) Yuan

About

I want to build powerful, human-centered LLMs — reliable enough to trust and wired to real-world resources — so that AI frees people's productivity and creativity, and makes everyday life genuinely better.

I'm an honors graduate of the University of Southern California's M.S. in Computer Engineering and a Shanghai Jiao Tong University alumnus, researching at USC's Fortis Lab, advised by Yue Zhao. My research spans agent safety, LLM reasoning, and world models — the pieces I believe a trustworthy, human-centered AI needs — and I turn each idea into a shipped, open-source system.

The through-line is reliability: an AI you can hand real resources to must verify its own actions, reason soundly, and model the world before it acts. Along the way I've built recommendation models at JD.com, low-latency trading infrastructure at Optiver, and autonomous GEO/SEO agents at Topify.AI. Outside research I'm building Serica, turning global curiosity about China into real trips.

Education

University of Southern California

M.S. in Computer Engineering, Honors Program · Fortis Lab
2025 – 2027Los Angeles, USA

Shanghai Jiao Tong University

B.Eng. in Electrical & Computer Engineering · Minor in Computer Science
2022 – 2026Shanghai, China

HSFZ · CCNU No.1 High School 华中师大一附中

High School Diploma
2019 – 2022Wuhan, China

Currently Building

CUA-frontier

A benchmark for computer-use agents beyond web and office apps — games, physical and embodied AI, chip design and CAD — graded on what the agent actually produces. Core contributor; organized by Prince Zizhuang Wang (CMU) with USC, UCLA, ASU, UW–Madison, Amazon and Scale AI.

Team ↗ Site ↗

AEGIS

The runtime safety layer for AI agents — cryptographic audit trail, human-in-the-loop approvals, and a kill switch, adopted with zero code changes. Open-source, live at aegistraces.com.

52656 GitHub ↗ Site ↗ USC Viterbi News ↗

Sovereign-OS

Charter-governed AI orchestration: a YAML constitution sets mission, budget & rules; CEO plans → CFO approves → Ledger tracks every cent & token → Auditor scores. 16 workers, Stripe, MCP. Open-source.

10514 GitHub ↗ Site ↗

Serica

A China-inbound travel platform that turns global curiosity about China into real, bookable trips — AI-generated itineraries, cultural adaptation, and a conversion funnel built for overseas travelers.

goserica.com ↗

BrewNet

An AI agent that makes professional warm introductions — it reasons over your network and goals to broker curated, high-signal intros to the right people, then handles the outreach end to end.

115 GitHub ↗ Site ↗

Publications

2026 · first-author unless noted. Aojie Yuan highlighted. See Google Scholar for the full list & citations · PDFs & BibTeX · glossary.

AEGIS architecture — Figure 1

AEGIS: No Tool Call Left Unchecked — A Pre-Execution Firewall and Audit Layer for AI Agents

A. Yuan, Z. Su, Y. Zhao

arXiv:2603.12621 Flagship system arXivdemoexplainer
WeClawArena — auditable cross-user agent sandbox

WeClawArena: An Auditable Sandbox and Benchmark for Cross-User Agent Collaboration and Security in Human-Centered Agent Networks

P. Z. Wang, A. Yuan, H. Zhang, X. Hu, Y. Zhao, S. Jiang

EMNLP 2026 arXiv:2608.03499 arXivPDFexplainer
CUA-SWE — benchmark overview and frontier-agent performance

CUA-SWE: When Computer-Use Agents Meet Visual Software Engineering

P. Z. Wang, C. Liang, Z. Xu, A. Yuan, X. Zhou, H. Zhang, Y. Zhao, X. Hu, S. Jiang

Under review · ICLR 2027 arXiv:2609.32600 arXivPDFexplainer
SWE-Team-Bench pipeline — a SWE-bench instance becomes a team task over personal workspaces, scored on task, workflow, and constraint success

SWE-Team-Bench: Benchmarking Coding-Agent Teams over Personal Workspaces

P. Z. Wang, A. Yuan, S. Jiang

Under review · NAACL
Auditable Agents — figure

Auditable Agents

Y. Nian, A. Yuan, H. Zhang, J. Li, L. Li, X. Hu, H. Wei, X. Xiao, C. Xiao, Y. Zhao

KnowFM @ ACL 2026 ACM AI Summit 2026 arXiv:2604.05485 arXivPDFexplainer
When Simulation Lies: A Sim-to-Real Benc — figure

When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents

X. Zhou, A. Yuan, Z. Luo, Z. Ling, X. Pan, Y. Gao, H. Zhang, J. Li, S. Jiang, et al.

NeurIPS 2026 FAGEN @ ICML 2026 arXiv:2605.11928 arXivPDFexplainer
SEVA: Self-Evolving Verification Agent w — figure

SEVA: Self-Evolving Verification Agent with Process Reward for Fact Attribution

A. Yuan, Y. Nian, H. Zhang, Z. Su, Y. Zhao

AI4GOOD @ ICML 2026 FAGEN @ ICML 2026 arXiv:2606.29713 arXivPDFexplainer
AgentIR — architecture

AgentIR: A Workload-Adaptive Cascade Retrieval Substrate for Long-Term Conversational Memory

A. Yuan, H. Zhang, S. Nazarian

arXiv:2605.25092 arXivPDFexplainer
Not All Thoughts Need HBM — architecture

Not All Thoughts Need HBM: Semantics-Aware Memory Hierarchy for LLM Reasoning

A. Yuan, T. Shen, D. Zhang

arXiv:2605.09490 arXivPDFexplainer
Sovereign-OS — architecture

Sovereign-OS: A Charter-Governed Operating System for Autonomous AI Agents with Verifiable Fiscal Discipline

A. Yuan, H. Zhang, Z. Wang, Y. Zhao

arXiv:2603.14011 arXivPDFcodeexplainer
Hidden Error Awareness — architecture

Hidden Error Awareness in Chain-of-Thought Reasoning: The Signal Is Diagnostic, Not Causal

A. Yuan, Z. J. Su, H. Zhang, Y. Nian, Y. Zhao

MechInterp @ ICML 2026 arXiv:2605.09502 arXivPDFexplainer
Beyond Language / FARS — architecture

Beyond Language: Format-Agnostic Reasoning Subspaces in Large Language Models

A. Yuan, Z. Su

arXiv:2605.09496 arXivPDFexplainer
Concept Subspaces — from controlled concepts to readout geometry

Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of Readout

A. Yuan, Z. J. Su, H. Zhang, Z. Su

arXiv:2609.39263 arXivPDFexplainer
Jacobian Rank Collapse in Decision-Focused Learning — figure

Jacobian Rank Collapse in Decision-Focused Learning

A. Yuan, H. Zhang, Z. Su

arXiv:2609.39261 arXivPDFexplainer
Agent Chronoception — the three proposed tests: preserve or switch, still valid at dispatch, is replanning worth it

Agent Chronoception: Temporal Validity and Useful Persistence

A. Yuan, Z. Su

Preprint · Zenodo DOIPDFexplainer
Auditing Agent Handoffs — metadata-only audit path, separate local release gate, and the three evaluation regimes

Auditing Agent Handoffs Without Centralizing Raw Messages: Evidence Boundaries and Release Tradeoffs

A. Yuan

Preprint · Preprints.org DOIPDFexplainer
HMAR — architecture

HMAR: Hierarchical Modality-Aware Expert and Dynamic Routing Medical Image Retrieval Architecture

A. Yuan

arXiv:2603.16679 arXivPDFexplainer
Do Agent Skills — cross-lingual security vs behavioral rates

Do Agent Skills Speak Safety in Every Language? A Cross-Lingual Security Analysis of the Skills Ecosystem

H. Zhang, A. Yuan, Y. Nian, Y. Zhao

Agent Skills @ ACM CAIS 2026 OpenReviewexplainer

Experience

Topify.AI  Forward Deployed Engineer

Nov 2025 – Jan 2026
  • Designed & deployed an autonomous GEO/SEO agent system (LangGraph, Inngest, Claude API) orchestrating 6 specialized sub-agents — keyword analyst, content strategist, writer, publisher, visibility monitor, strategy optimizer.
  • Built a closed-loop pipeline: the monitor tracks brand share across AI search engines (Perplexity, ChatGPT, Gemini) and feeds visibility deltas back to the optimizer, which dynamically tunes content structure, topic selection & publishing cadence.
  • Cut content-ops headcount by 80% and drove 200% revenue growth to $3M ARR, replacing a 5-person team with a single self-optimizing workflow.
LangGraphInngestClaude APIMulti-agent

JD.com  Recommendation Algorithm Engineer · Beijing

Dec 2024 – Feb 2025
  • Helped build an end-to-end generative recommendation framework (semantic-ID retrieval), unifying candidate generation and ranking into one sequence-to-sequence model replacing the multi-stage cascade.
  • Used NER to mine 18 key attributes from year-long logs; applied Qwen-VL for item visual-text analysis and Aho-Corasick for efficient query matching, strengthening semantic features and improving codebook utilization for ID quantization.
  • Built a multi-view user-behavior module fusing short/long-term interaction sequences via QFormer-based aggregation of hierarchical semantic embeddings from click, order & RSU behaviors.
Generative RecQwen-VLQFormerSemantic ID

Optiver  Algorithm Engineer · Shanghai

Dec 2023 – Feb 2024
  • Developed low-latency components for a volatility-prediction pipeline targeting Chinese commodity futures, supporting automated market-making.
  • Accelerated an LSTM volatility model by porting the engine from Python to C++/LibTorch with custom CUDA kernels; achieved a 70% inference-latency reduction (5 ms → 1.5 ms) via TensorRT quantization & kernel fusion.
  • Built the real-time market-data ingestion module in C++/ZeroMQ to process tick-level order-book data from Chinese futures exchanges.
C++LibTorchCUDATensorRTZeroMQ

Enjoymi Games 欢动科技  Software Engineer · Shanghai

Aug 2023 – Oct 2023
  • Maintained core SDKs for two major mobile games — robust login, payment & platform integration.
  • Engineered custom SDK integrations for major Android channels (e.g. Xiaomi channel package) and architected a unified authentication interface over the HMS Account SDK, exposing a clean login() handling silent/explicit sign-in and all auth states.
AndroidJavaHMSSDK

Awards & Honors

  • — AEGIS named among the funded open-source projects (USC Fortis Lab) 2026
  • — AI auditing & agent security; AEGIS & Sovereign-OS among the named projects (USC Fortis Lab)
  • — Public Value Award · Evermind Hackathon 2026
  • National Scholarship 国家奖学金 — China's national undergraduate scholarship, awarded to ~0.2% of students 2023
  • — Gold Award 2024
  • — Silver Award 2023
  • — Second Prize
  • The University Physics Competition (UPC) — Silver Award 2023
  • — Honorable Mention 2023

Academic Service

  • — International Conference on Learning Representations (ICLR) 2027
  • — International Conference on Machine Learning (ICML) 2026
  • — Conference on Empirical Methods in Natural Language Processing (EMNLP) 2026