Hello, I'm Alvin

AI Engineer · Developer Tools · Full-Stack

Building independent AI products and developer tools. Scroll for selected work, or jump to Projects.

Scroll to explore
Profile

Alvin Xu

Software engineer — AI agent systems, retrieval, developer tools

yingjiex@usc.edu
linkedin.com/in/yingjie-xu-435a111b5
Vancouver, BC · Canada-remote
Open to AI Engineer / Founding Engineer roles
University of Southern California — M.S. in Computer Science
Aug 2023 — May 2025 · Los Angeles, CA
McGill University — B.Sc. (Honours) in Computer Science
Sep 2019 — May 2023 · Montréal, QC
LLM agent loops (Plan-Execute, ReAct)Two-stage retrieval (BGE reranker + RRF fusion)Multi-LLM orchestration & cost-aware executionMCP-compatible inspection / automation serversTauri + Rust + React desktop appsLayered Rust DB architecture & E2E test harnesses
AI / ML
LLM agentsRAG / retrievalrerankingembeddingsmulti-LLMMCPPyTorchNLP
Languages
PythonRustTypeScriptC/C++JavaSQLBash
Web & App
TauriReactNext.jsNode.jsExpressPixiJSAngularTailwind
Data / DevOps
PostgreSQLMongoDBSQLiteAWSAzureDockerGitHub ActionsGit
Robotics
ROSRRTMCLSLAMIK / TSR
Focus Now
  • CiteLoom v1: agent-driven research canvas — Plan-Execute v3 loop, two-stage BGE + RRF retrieval, MCP-compatible inspection server
  • NeuralLens Phase 1: mechanistic interpretability primitives for VLA / world models / RL policies, with live simulation-loop hooks
  • Internal A/B benchmarks: ranking evaluation (NDCG / MRR), agent cost-aware execution, run-to-run variance reduction
  • Solo build discipline: 29 ADRs, A/B-validated subsystem migrations, pixel-level E2E test harness over Chrome DevTools Protocol
Strengths
  • LLM agent design: Plan-Execute and ReAct patterns, cost-ceiling reasoning, multi-provider abstraction
  • Retrieval engineering: BGE cross-encoder reranking, RRF fusion, embeddings, graded-query evaluation
  • Systems delivery: Tauri + Rust + React desktop apps, 4-layer Rust DB architectures, MCP automation servers
  • Full-stack range: Next.js / Node / Mongo / Azure, Kotlin Android companions, ROS robotics fundamentals
Principles
  • A/B everything: pick the winner by metric, not intuition — NDCG, MRR, picks, run-to-run variance
  • Document decisions: ADRs over heroic memory; future-you cannot debug yesterday-you
  • Diagnose before patching: root-cause the bottleneck (e.g. cost-ceiling truncation), not the symptom
  • Honest engineering: real-pixel tests, internally graded benchmarks, no production-scale claims without users
Selected Work
CiteLoom — Independent AI Product
LLM AgentRetrievalMCPTauriRust

Tauri-based AI research canvas combining a Plan-Execute LLM agent loop and a two-stage BGE + RRF retrieval pipeline. 29 ADRs, an MCP-compatible inspection/automation server (163 endpoints across 6 namespaces), and a 4-layer Rust DB architecture.

NeuralLens — Mechanistic Interpretability Toolkit
PyTorchMech InterpRoboticsOpen-source

Open-source Python library (in development) for inspecting and steering robot foundation models — VLA models, world models, RL policies — with activation inspection, feature steering, ablation, and live simulation-loop hooks.

Investment Platform — Full-stack Web & Android
AngularNode.jsMongoDBKotlin

Stock dashboard (Angular, Node.js, MongoDB, Azure) integrating Finnhub and Polygon APIs for live market data + Highcharts visualizations, with a Kotlin Android companion app sharing the same backend.

6-DOF Robotic Arm — Motion Planning & Control
ROSRRTMCLLiDARTSR-IK

Motion planning + localization framework (RRT trajectory planning, MCL with LiDAR) and grasping (TSR-based IK sampling + Jacobian methods), validated on both RViz simulation and a real robotic arm.

End-to-End NLP Pipeline (Classical ML → Transformers)
BiLSTMTransformerGloVeNLP

Sentiment, POS, and NER pipelines on Amazon Reviews and CoNLL-2003. Benchmarked classical ML (SVM, HMM, LR) against neural models (Word2Vec, GloVe, BiLSTM, GRU, Transformer); BiLSTM+GloVe reached 88% F1 on NER.