From requirements to architecture
I turn product requirements into clear service boundaries, data ownership, and failure-handling decisions.
Hi, I'm
Xiaonan Li
AI Systems & Full-Stack Engineer
I build AI features and the systems that run them.
I'm completing a Master of Computer Science at the University of Sydney, focused on data science and AI. I like turning a rough requirement into software that is reliable and maintainable.
Jump to2026
Looking for teams building reliable AI products.
How I work
I turn requirements into software that is easier to build, test, and operate. That includes the feature itself, but also service boundaries, data flow, and failure handling.
Based in Sydney, Australia.
Long-running AI workflows
I design AI workflows with asynchronous workers, persisted progress, retries, and clear boundaries from the core backend.
Across the stack
I have worked across backend, frontend, AI integration, and deployment to take a feature from design through delivery.
Working languages
Chinese is my native language. I study and work in English, and I am continuing to learn Japanese.
Engineering internships
Experience in production troubleshooting, reliability improvements, testing, data analysis, and developer tooling.
Jun 2025 - Aug 2025
Meituan Platform
Software Development Engineer Intern
Worked on reliability and developer productivity in core systems, including resilient status reporting, issue diagnosis, data analysis, and Gradle build optimisation.
- 01
Improved status reporting and retry logic to increase reliability under weak network conditions.
- 02
Used JADX to diagnose an obfuscated third-party reporting-format defect and correct the payload format.
- 03
Built Spark and SQL dashboards for anomaly detection and trend analysis.
- 04
Analysed the Gradle task DAG and optimised build bottlenecks while preserving build stability.
- 05
Used AI-assisted tools for unfamiliar code analysis, API exploration, and issue investigation.
Education
Two undergraduate degrees in China and Australia, followed by a master's in Sydney focused on data science and AI.
Scroll sideways
Now — graduating Nov 2027
The University of Sydney
Master of Computer ScienceData Science and Artificial Intelligence
Current master's study in machine learning, data pipelines, and rigorous evaluation of model results.
2026 — graduated with Distinction
Swinburne University of Technology
Bachelor of Engineering (Honours)Software Engineering
Honours study in software engineering, including AI, applied machine learning, and team software delivery.
Graduated
Shandong University of Science and Technology
Bachelor of EngineeringSoftware Engineering
Undergraduate study in computer science fundamentals. My thesis examined federated learning across devices that cannot share their data.
Selected projects
Projects in AI security, AI platforms, federated learning, and computer vision.
Creator and Researcher / Jul 2026 - Present
RedCell — Adaptive AI Agent Red Teaming
An open-source tool that probes tool-using LLM agents for prompt injection, data leakage, and unauthorized tool calls within a fixed test budget, then turns confirmed findings into regression tests.
PythonPydanticSQLAlchemyThompson Sampling
View sourceChallenge
Tool-using LLM agents call databases, files, and business APIs, so a manipulated agent can leak data or take actions it shouldn't. A fixed list of jailbreak prompts doesn't cover this, because the attack surface comes from natural-language decisions and non-deterministic models.
Approach
I modelled red teaming as a budget-constrained search problem: a bandit controller decides which attack strategy gets the next attempt, and an LLM mutates the prompt. Success is judged by instrumentation — canaries, tool-permission checks, cross-user access — not by another LLM.
Implementation
Built the full pipeline with a deliberately vulnerable benchmark arena, three controllers, budget and reliability guards, experiment fingerprints, and crash-safe resume. The Phase 0 hypothesis was not supported across 18 runs and 1,080 attempts, so I recorded it as a negative result.
Tech stack
Python 3.11, asyncio, Pydantic, SQLAlchemy, Typer CLI, Jinja2 reports, pytest. The test suite runs fully offline.
- 01
Designed adaptive attack search: bandit-guided strategy selection with LLM-based prompt mutation under a fixed attempt, token, and cost budget.
- 02
Built deterministic scoring from canary strings, tool-permission checks, and cross-user access, separating violating intent, attempted tool calls, and realized impact.
- 03
Built a benchmark arena of deliberately vulnerable agents, with positive and negative controls plus a check that validates the attacker model itself before any calibration run.
- 04
Made every run reproducible with experiment fingerprints, full trace storage in SQLite, crash-safe resume, and export of confirmed findings as regression tests.
- 05
Reported the Phase 0 research hypothesis as NOT SUPPORTED after 18 runs and 1,080 attempts, and documented the evidence limits publicly rather than reporting a favourable subset.
Technologies I use
Technologies I have used in projects and internships.
AI & Machine Learning
Used for model training, video liveness detection, semantic retrieval, and LLM agent evaluation.
PyTorchTensorFlowTensorFlow LiteRAGFAISSSentenceTransformers
Backend & Data
Used to build APIs, manage application data, and deploy services.
Spring BootJavaMyBatisMicroservicesPythonPydanticSQLAlchemySparkSQLAWS
Frontend & Mobile
Used to build web and Android interfaces, including this site.
Next.jsTypeScriptTailwind CSSAndroidKotlinJUnit
Get in touch
I'm looking for internship and graduate engineering opportunities in AI systems, backend, and full-stack development. The form, LinkedIn, and GitHub are all good ways to reach me.
Currently seeking
Internships and graduate roles
Opportunities in AI engineering, backend systems, and full-stack development.