Yizhou (Russell) Lu · Software Engineer — AI Applications & Full Stack
My work spans LLM-powered products, data-quality evaluation, and the full-stack systems around them. I focus on making results explainable, failures recoverable, and releases verifiable.
Flagship — JobSignal
Resume in, ranked visa-friendly jobs out. Claude extracts a candidate profile, pgvector retrieves listings sourced from public Department of Labor filings, and an LLM reranker writes a rationale for every match.
Live Ingestion repair merged · CI passing Production acceptance pending
Software Engineer
"Strong fit: pgvector + LLM pipeline work maps to your retrieval stack; H-1B sponsorship history confirmed in DOL filings."
Two more pieces of work with a written case each: a development tool I built and use, and my internship engineering.
Developer tooling
Task routing, independent review, and verifiable delivery
Built a reusable local CLI for routing development tasks across LLM providers, with isolated workspaces and traceable run evidence. Used it within a coordinated workflow for a JobSignal repair pilot and the LLM Interview Lab delivery, keeping execution success separate from acceptance.
Local CLI preview Cross-model routing Independent review
Internship engineering
Release verification for a computer-vision product
Contributed release-verification tooling to FreshLens during my TalkMeUp internship, including deterministic gate orchestration, shared result contracts, and reproducible evidence for review and handoff. Earlier in the internship I worked on YOLOv8 produce detection.
Orchestrator merged Some materials in review
An evaluation case, a full-stack SaaS, and the public half of my delivery process.
A conservative LLM check that vetoes a price match when the product identity does not line up: wrong model, accessory, bundle, or another item that should not be mixed in. Deterministic rules handle the clear-cut conditions; the model only reviews semantics and never generates a price. Measured false rejections, missed mismatches, repeatability, latency, and cost separately, before any production integration.
Offline evaluationNot in production
team system · my scope: the veto layer and its evaluation
Evidence-backed delivery tracking. Links GitHub events to work-item state so progress is grounded in what actually happened: deterministic rules draw the conclusions, AI explains them. Full-stack SaaS built around OAuth, webhooks, native work items, and background processing. A preview is deployed, but the product is not fully functional yet; the source shows the current state.
In progressPreview deployed · not fully functional
Role templates for implementation and independent review, delivery rules, CI templates, and lessons learned from running AI-assisted delivery. The public, reusable templates behind the workflow above; the CLI itself is a separate private preview.
Public repository
github.com/RussCTGL/agent-harness-pack ↗Where I turn what I'm studying into tools I can maintain.
A self-directed AI fundamentals learning app with 150 practice questions, 98 knowledge notes across 16 modules, strict multi-select grading, and in-browser progress recovery. Built with content validation and independent review before release. Content ships with the site; it does not call an LLM at runtime.
llm-interview-lab.vercel.app ↗Bilingual notes for USC CS544 (Applied NLP), organized by topic and linking concepts, worked examples, formulas, and sources. Follows the course lectures and credits them; not an original curriculum.
Team course project on product catalog matching: problem framing, abstention with human review, and experiment design. No trained results yet; baseline and error analysis to follow.
In progressTeam project
The short version. The full version is on my resume.
FreshLens scores produce freshness with computer vision. I started on YOLOv8 detection and finished owning release verification for the team.
Gate orchestrator mergedSome materials in review
Case study →Built a full-stack e-commerce trading platform (Django, React, MySQL) with authentication, listings, and transactions; designed its REST APIs and moved it from on-prem to AWS (EC2, RDS) with CI/CD.
Built an AI bot that automated Google Sheets workflows, replacing roughly two weeks of manual customer analysis with a few hours of unattended batch processing.
MS in Computer Science, University of Southern California, Aug 2025 – May 2027. Coursework includes Applied NLP (CS544).
BS in Computer Sciences and BS in Mathematics, University of Wisconsin–Madison, 2022 – 2025.
Much of my recent work is built with AI coding tools across several models, so my job is defining the requirement, reviewing the evidence, and deciding what ships. One recent run exited successfully, but its run evidence did not cover the acceptance criteria. It went to awaiting verification instead of done, and was reworked. A successful exit is not acceptance.
MS CS at USC, Class of 2027. Seeking AI / LLM engineering and full-stack roles where model output has to be explainable and releases have to be verifiable.
0109yizhoulu@gmail.com