Justin Xingyuan Zhao
I build LLM systems and agentic workflows. Building them made me obsessed with one question: when should you trust them. The interesting problems live in the gap between what a model scores on a benchmark and what it actually does when conditions shift.
The pipelines I have run handling hundreds of millions of interactions rely on vLLM and llama.cpp for deployment, Kubernetes for load balancing, LoRA and GRPO for fine-tuning, and LangGraph and customized framework to orchestrate workflow: running them at that scale is where I stopped trusting benchmarks. The field's most urgent need is not better models, but better epistemics around models, the tools and methods to know what a system actually understands, to measure trust rather than assume it, and to fail predictably rather than mysteriously. That is the problem I am working on.
Timeline
From a price monitor script in 2019 to production LLM systems and PhD research today, each phase added a layer: industry rigor, startup velocity, research methodology, and engineering depth.
Democratizing LLM Development: NoCode Platform
Architected a production NoCode LLM application using React Flow for visual orchestration and a FastAPI/Celery backend for asynchronous multi-agent execution across multiple OpenRouter endpoints. Users build complex LLM pipelines through drag-and-drop, reducing weeks of onboarding to minutes of configuration.
Price Monitor 2.0: Agentic Trading System
Built an agentic trading framework where agents fetch live market data, build context autonomously, generate decisions, and review each other's outputs before acting. Customizable orchestration layers bring context-aware, auditable trading to non-technical users without engineering overhead.
Investing in the Future: Teaching and Mentorship
Taught and mentored across academic programs and research communities, bringing quantitative methods and technical skills to researchers at multiple career stages. Shaped research workflows and technical foundations for 200+ junior researchers and students, translating graduate-level methodology into actionable practice.
M.S. Computer Science and SDAR Program
Joined the M.S. Computer Science program at the University of Colorado Boulder, focused on machine learning and cloud computing. Also joined the Social Data Analytics and Research (SDAR) program at UT Dallas in 2023 alongside the PhD. The combination of production engineering training and rigorous statistical methods is the foundation that connects every system and paper above.
Research Journey: Ph.D. and Research Assistantship
As Research Assistant at UT Dallas, builds research infrastructure for large-scale NLP and LLM work, handling noisy real-world corpora at a scale that demands production-grade engineering. The dissertation establishes principled criteria for when LLMs can and cannot be trusted for social science research tasks, giving researchers evidence-based guidance rather than trial-and-error. Current work spans noisy corpus pipelines, agentic RAG for policy analysis, LLM bias evaluation, and multimodal system comparisons on real-world image detection tasks.
Knowledge, Released: Research and Conferences
Began building research at the intersection of computational methods and social science, developing skills in text analysis, causal inference, and applied LLM pipelines. This phase established the methodological foundation that the conference papers and dissertation work built on.
Public Relations Analyst, ByteDance
Applied analytical rigor to advertising strategy across multiple product lines, consistently driving traffic growth through data-informed decisions. Brought an automation mindset to reporting and campaign analytics, shifting the team from manual data collection to strategic work.
Builder Journey: AI Products and Ventures
Started building AI products in 2019 and has continued since. As founding engineer at Prompter Store LLC, shipped three concurrent consumer products from ideation to production: a multi-modal AI chat platform that scaled to hundreds of millions of user interactions, an AI tools directory, and a consumer network product.
Where It Began: Price Monitor, 2019
Built the first iteration of Price Monitor in 2019, a rule-based system for tracking prices and surfacing actionable signals. The same core problem that drove the 2019 version, making market information legible and actionable without constant manual monitoring, is what the 2026 agentic system solves at a fundamentally different level of autonomy.