Systems Builder and Researcher
Hui Xu at Stony Brook University. RL, systems, and AI research.
I build and ship scalable products across global teams. I have owned and operated 10+ Linux cloud servers end-to-end (OCR pipelines, invoice verification services, databases, deployment). Algorithmic improvements reduced manual review while improving recognition speed and accuracy. I am currently focused on reinforcement learning for RTS environments and robust agent evaluation.
Goal
Build great products that create value for individuals and enterprises, enabling shared wins.
I have led full-stack operations for startups, maintaining more than 10 Linux cloud servers including OCR, invoice validation, and database systems. I implemented algorithmic optimizations that reduced human QA cycles while speeding up recognition with higher accuracy.
In multinational environments, I communicate across teams and stakeholders in the US, India, and China to deliver products on schedule and with clear ownership. Delivered on-prem deployment of a full-stack business analytics platform.
Latest milestones, RL progress, and research highlights.
Project update
I am working on QuantDesk to help me research stocks and make more informed investment decisions. It brings together price trends, support and resistance levels, volume, and momentum in a visual dashboard.
Career update
Recently completed my internship at Amazon, gaining hands-on experience building and delivering software at scale.
Competition
Our team ranked 1st on the overall private leaderboard in the AgentX-AgentBeats Agent Safety track, hosted by Berkeley RDI and sponsored by Lambda.
Leaderboard CompetitionPaper acceptance
Our paper (co-first authors: Hui Xu and Xiaochuang Yuan) was accepted by the ICLR 2026 AIWILD Workshop.
OpenReviewAcademic service
Serving as a reviewer for agent-related submissions. Workshop date: April 26, 2026 (Rio de Janeiro, Brazil).
WorkshopCompetition
Our team placed 2nd in the AgentX AgentBeats finance agent competition organized by UC Berkeley RDI and sponsored by OpenAI.
Details Our StoryResearch update
Adaptive questioning for multi-condition mental health assessment.
ACL AnthologyApplied systems, research, and reliable AI infrastructure.
I am working on MAQuA as my thesis, exploring adaptive questioning for multi-condition mental health assessment with a minimal number of questions.
Thesis in progress; accepted to EACL 2026 (Main Conference Poster).
PaperI am learning parallel programming through Aalto University’s Open 2026 course, Programming Parallel Computers.
Course pageOptimizing stateful neural networks for streaming market data.
I optimize a PyTorch ensemble of Mamba2, RetNet, Hawk, and xLSTM models that maintains a separate recurrent state for each symbol. My implementation combines batching that preserves per-symbol request order, CUDA graph replay, mixed-precision linear layers, and fused Triton kernels for recurrent state updates.
This is part of my thesis. MAQuA enables assessment of multiple mental health conditions with a minimal number of adaptive questions. Current language models do not model the human subconscious. I am researching how to represent it, including L'inconscient inculqué à mon ordinateur (Book, French Edition) and Lacan and subjective topology.
Accepted to EACL 2026 (Main Conference Poster). An additional NLP research project is ongoing.
PaperWhat I cannot create, I do not understand.
Repo700+ problems solved. Preparing video walkthroughs for challenging problems to build skill in reading, verifying, and improving LLM-written code for correctness, efficiency, and safety.
I have implemented three distributed transaction systems in both Go and C++23: stable-leader Multi-Paxos, PBFT with signatures and view changes, and sharded Multi-Paxos with two-phase commit. The projects cover fault handling, recovery, checkpoints, benchmarking, and resharding, with shared test cases to compare behavior across the two languages. I also plan to read the underlying papers and record tutorials.
RepoResearch outputs and works in progress.
Accepted to EACL 2026 (Main Conference). Preprint on adaptive questioning for multi-condition mental health assessment.
ACL AnthologyPeer review and community contribution.
Reviewing submissions on trustworthy, safe, and effective agent systems for the workshop held on April 26, 2026 in Rio de Janeiro.
Workshop SitePaintings and sketches by me.
I am a lesbian; I value the Statue of Liberty's ideals of freedom, democracy, hope, and inclusion, alongside bold exploration backed by safety.
Personal essays on Lacanian psychoanalysis (Chinese).
Earlier projects that shaped my product and engineering craft.