Minzu University of China2023 — 2027
B.S. Computer Science and Technology
Coursework: data structures, operating systems, computer networks, computer organization, databases, software engineering, natural language processing. GPA 3.6.
Class of '27 · CS @ MUCAgent / RAG / LLM systems
Sean Yu
SEAN YUAI-native developer, combining product taste with engineering discipline to turn LLM ideas into traceable, evaluable systems.
SEAN YU
Agent / RAG / LLM systems
“ I judge LLM applications on reproducible evals and end-to-end traces, not on gut feel. ”
Agent lab notes
From prompt to trace to measurable behavior.
Trace
tool calls
structured inputs, observable failures, retry paths
Retrieval
hybrid RAG
embedding + lexical search, reranking, context control
Eval
rubrics
case suites, score deltas, reproducible regressions
Scroll
about
AI Native developer focused on traceable Agent, RAG, and workflow systems. My default loop is prototype -> trace -> evaluate -> harden: build quickly, expose tool calls and context, measure behavior with evals, then stabilize with Docker, GitHub Actions, and reviewable workflows. I also treat my LLM Wiki as a Content Operating System for turning project work into reusable knowledge assets.
skills
Languages & backend
PythonTypeScriptFastAPIPydanticpytestAsync I/O / asyncioMySQLRedis
Frontend & streaming
Next.jsReactVercel AI SDKStreaming UITailwind CSS
LLM & agents
OpenAI APImulti-model adapterstool callingstructured outputAgent Loop Design (ReAct, tool orchestration)Context Engineering
RAG & retrieval
embeddingsQdrantChromahybrid retrievalBM25 / RRFrerankersRAGAS-style eval
Engineering
DockerGitHub Actionslogging & error handlingVercelremote collaboration
projects
shotgunCV
Pipeline-first AI Resume Ops for bulk applications: parse JDs, generate resume variants, score and rank them, then output application strategy. It treats resume tailoring as an evaluable workflow, not a one-shot prompt wrapper.
- Problem
- Bulk applications turn resume tailoring into a comparison problem: many JDs, many variants, and little objective signal for priority.
- Architecture
- JD parsing -> resume variant generation -> scoring/ranking loop -> strategy output, with quality checks before scoring.
- Evidence
- Grounded in the knowledge-base workflow notes: parse -> generate -> score -> rank -> strategy makes each stage independently eval-oriented.
PythonLLM APIPydanticPipeline orchestration40GitHubMergeWarden
AI Code Review Agent with structured output, a ReAct tool-execution loop, and golden-set evaluation. Hard timeout plus budget soft cap cut worst-case single model-call latency 84% (375s -> 60s), while diff-first 80-line windowing cut per-file prefetch context 95.1% (75,466 -> 3,691 chars).
- Problem
- Code review agents fail when they overfill context, stop with summary-only output, or turn timeout paths into fabricated findings.
- Architecture
- ReAct agent loop: QueryEngine -> Orchestrator -> Tool System -> Force-Submit -> Structured Review, with Pydantic validation and DeepSeek v4-pro thinking-mode protocol handling.
- Evidence
- Golden eval recovered from 0% to 50% hit rate, eliminated false positives 100% (3 -> 0), and held schema validity at 100% through force-submit closure, severity floor, and expected-location relaxed confidence.
PythonFastAPIDeepSeek v4-proPydanticDockerpytestasyncio10GitHub
activity
671 contributions in the last 6 months
DecJanFebMarAprMayJun
Mon
Wed
Fri
LessMore
education
languages
English
IELTS 7.0 · CET-6Chinese
Nativehighlights
Raw PRDiff, CI signal, repo context
MergeWarden
Tool Design
Tools are permission-tiered as readonly / write / execute, with allowlists, shell=False, environment cleanup, output truncation, and high-risk gating to enforce the Agent action boundary.Safe merged PREvidence checked, schema valid, merge ready