Find the real problem
At Shopee, the bottleneck was not query accuracy. PMs without a SQL background could not write precise data requests or find an existing, mature Topic. I turned that gap into an end-to-end application problem.
AI Full-Stack Engineer · Agent Applications & Data Systems
From business problem to reliable AI application.
How I work
Automate execution, never understanding
People may hand execution to AI, but they must still see why. Aesthetics, evaluation, and decisions stay human; uncertain cases go back to a person.
Constrain what must not fail
SQL templates, output schemas, lint rules, explicit failure modes — deterministic code guarantees what the model cannot.
Analysis Agent Copilot
SvelteFrontend
FastAPILayered backend
69Version-bound cases
Retrieve → Select → Authorize → Generate · Allowlisted Tool Use
I owned the full cycle from requirements and architecture to delivery and iteration across a Svelte frontend, layered FastAPI backend, Python ingestion pipeline, and Git-versioned storage. The application is in daily analyst use; retrieval evaluation drove adoption of Pi as the orchestration layer, now completing performance tuning before rollout.
A/B Test Agent Skills
Reusable Agent Skills for recurring A/B test analysis
40+ UAT questions · Accuracy target met · In business use
ContextCue
The model sees only the gate-filtered memory subset
Public live prototype · 16 synthetic scenarios
Projects
Runnable systems
02ContextCue
A governed Agent memory layer that isolates memories by person and task scope, then filters them by evidence, privacy, and expiry.
16 synthetic scenarios · Scoped memory · TTL · Privacy
Clinical Bazi Web
A public full-stack app combining code-based chart calculation, true-solar-time correction, and streaming AI-assisted interpretation.
Public · Runnable end to end
Applied agents, ML & data systems
05Hand-Coded Adversarial Agent
A pure-Python battle-royale agent designed through black-box analysis and ablation—not neural-network training.
Tied #1 / 6 · 43% vs 70M-step NN
Sim2Real Calorie Prediction
A physics-informed data pipeline that closes the gap between gym training data and real daily activity.
External R² −14.38 → 0.79
Mini-GPT
A decoder-only Transformer built from scratch in PyTorch, including causal attention and a training loop.
Transformer from scratch
EvoRed
An evolutionary red-team system that evolves adversarial prompts and includes a defense module.
Attack evolution + defense
Semi-Automated Data Labeling
A human-guided labeling pipeline that automates confident cases while preserving review for uncertain ones.
100% on evaluated set · 84% automation
Human–AI, learning & knowledge
05AI / Python Teaching Scaffolds
Curricula, active-learning activities, and reflection-oriented prompt rubrics for applied ML and GenAI.
Lecturer, 2023–2025 · 300+ students
Obsidian Knowledge-System Refactor
An agentic refactor of a multi-year knowledge base with inspection, rollback, and human review.
Hundreds of notes · Human review
Synapse
A concept prototype that generates notes, knowledge graphs, and quizzes from video lectures.
Notes · Knowledge graphs · Quizzes
Yuehuo AI Disability Assistance
An AI assistance concept grounded in user research and data-driven needs analysis for disabled users.
User research · Needs validation
AI Feedback Format & Durable Learning
A research sketch comparing direct explanation, Socratic scaffolding, and self-evaluation-first interfaces.
3 interface conditions · No results yet
Recognition
Honors
-
Competition · ProjectNational Finals · 3rd Prize
RAICOM Autonomous Indoor UAV
Aerial Reconnaissance · Project lead · ROS
-
ScholarshipTop 0.2% nationwide
China National Scholarship
Suzhou University · B.Eng.
-
Provincial recognition
Outstanding Graduate of Anhui Province
Anhui Province · China
-
Competition · ProjectGlobal Final · 3rd Prize · Top 5%
Huawei ICT Competition
Experience
The journey
B.Eng. Data Science & Big DataSuzhou University
China National Scholarship, placing in the top 0.2% nationwide in 2022–23.
LecturerJinken Vocational & Technical College
Taught AI and Python to 300+ students; designed applied-ML and GenAI curricula and active-learning activities.
M.Comp. in Applied AINanyang Technological University · CCDS
Multi-agent systems, trustworthy AI applications, and human–AI learning research.
AI Data Analyst InternShopee · Singapore
Delivered the missing link in an end-to-end PRD → data request → data Agent → result flow: one knowledge base unifying historical requests, business SQL, and Diana Topics, so PMs can produce structured, traceable requests in natural language and route them to the correct Topic.
Owned Analysis Agent Copilot from requirements and architecture through delivery and iteration: Svelte frontend, layered FastAPI backend, Python ingestion, Git-versioned storage, allowlisted Tool Use, and a persisted retrieve → select → authorize → generate state machine. It is in daily analyst use, backed by 69 version-bound historical cases and retrieval evaluation on Recall@5, nDCG@5, and citation accuracy; results drove adoption of Pi as the orchestration layer, now completing performance tuning before rollout.
Co-designed and executed the Data Agent Topic SOP, covering table selection, knowledge-base curation, a 40+ question UAT set, and error-taxonomy-driven regression fixes. The reusable A/B Test Agent Skills met the accuracy target and are used in real business analysis across multi-country BI teams.
Capabilities
Skills
- PythonBackend core
- TypeScriptFrontend core
- Svelte + FastAPIFull stack
- React + Next.jsWeb apps
- SQL + Data PipelinesData systems
- Tool Use + MemoryAgent applications
- Retrieval + EvaluationEvidence & quality
- Git + CIDelivery
About
I turn real business problems into reliable AI applications, building the frontend, backend, and data pipeline while constraining failure-critical steps with SQL templates, structured outputs, evidence, and explicit failure handling.