YL. Email me

Yizhou (Russell) Lu · Software Engineer · Los Angeles

Signal from noise, shipped with evidence.

I build AI products end to end — LLM retrieval pipelines, computer-vision scoring, the full stack around them — and I run fleets of AI coding agents under review gates strict enough that every merge carries proof. MS CS at USC, formerly UW–Madison (CS + Math).

GitHub ↗ LinkedIn ↗ JobSignal, live ↗
0+jobs ingested, 96% H1B-confirmed
0unit tests + Playwright E2E
0+PRs through review gates
0images hand-labeled for calibration
$1–2LLM cost per user / month

Work that shipped

Production systems and side projects, built in 2026. Cards with an arrow are live — click through.

Solo project Jul 2026 – present

JobSignal

An AI job-search platform for international candidates — live at liuzijob.com.

A resume goes in; a ranked universe of visa-friendly jobs comes out. Claude extracts the candidate profile, pgvector HNSW search retrieves matches from 8,400+ listings, and an LLM reranker writes a per-job rationale — rendered as an interactive 3D galaxy, at about $1–2 per user per month through per-user quotas.

  • Grew the job pool 250× (34 → 8,400+) with a pluggable ingestion system crawling Greenhouse, Lever, and Ashby boards, seeded from 27,428 real DOL filings and refreshed by daily crons.
  • Diagnosed a production latency breach down to an HNSW index defeated by a filter expression; shipped an exact-filter RPC that restored the plan.
  • Row-level-security multi-tenancy, prompt-injection hardening, and a dead-man's-switch daily health report — silence itself raises an alert.
  • 435 commits and 60+ PRs since July, every one through the agent review pipeline below.

Next.js · TypeScript · Supabase Postgres + pgvector · Claude API · Vercel · Playwright

Internship · TalkMeUp Jun 2026 – present

FreshLens

Computer-vision produce freshness scoring, on an 8-person agile team. The parts below are mine.

A camera looks at a produce shelf; the system detects each item, scores its freshness 0–100, and generates discount worklists and rescue recipes.

  • Multi-item YOLOv8 detection with produce-class filtering and confidence thresholds — 9 raw boxes down to 4 correctly labeled items, false positives eliminated; fixed a zero-shot label-override bug between classifier and detector.
  • Built the calibration ground truth by hand: 154 labeled images across 9 produce types (real photos and AI fakes), plus the proxy-calibration script and evidence-quality evaluator that scored the model against it.
  • Live detection feed in Next.js/React with dynamically scaled bounding-box overlays, used in stakeholder demos.
  • Designed the project's deterministic closeout gate orchestrator — frozen evidence packets, a v1 result contract, symlink refusal and env allowlisting — through two rounds of peer review.

FastAPI · YOLOv8 · Claude API · Next.js · pytest · GitHub Actions CI

Solo project Jun 2026 – present

GoThin

An AI fitness coach that makes logging effortless — live on Vercel.

Snap a photo of your meal or describe it in a sentence; GoThin turns it into structured nutrition data and tells you when you'll hit your goal.

  • Vision-based macro estimation with confidence scoring and one-tap undo; goal dates projected by linear regression over your actual logs.
  • LLM cost engineering: model routing, prompt caching (~7× cheaper input tokens), and schema-constrained structured outputs.
  • A trace dashboard tracks per-call cost, latency, and cache-hit rate — every API dollar is visible.

Next.js · TypeScript · Supabase Postgres · Claude API · Vercel

Game Jan – May 2026

Lumen Runner

A Unity 6 runner where light is ammunition — playable in the browser.

  • The map starts dark; shooting light projectiles is how you see where you're going, so every shot spends both ammo and safety.
  • Collectibles trade into instant reload and higher fire rate, turning pickup routes into a risk decision.

Unity 6 · C# · WebGL build on GitHub Pages

Twitch Gadget

Content-based stream recommendations from user favorites and game metadata, with Spring Security auth and Caffeine caching.

Spring Boot · React · MySQL · AWS

How I ship with agents

JobSignal is built by nine role-scoped AI agents I orchestrate — every issue walks this pipeline, and no agent approves its own work. Hover any station.

program-manager Roadmap → issues
Converts the roadmap into milestones, epics, and implementation-ready issues. Never writes code.
product-spec What, exactly
Translates product outcomes into precise user behavior and acceptance criteria on the issue.
architect How, exactly
Technical design, schema and interface contracts, ADRs. Multi-round designs end in a single design-of-record document.
implementer Code + tests + PR
Implements exactly within scope. Never approves, merges, or closes its own work — and never touches the hosted database.
review panel code · security · test
Independent line-level review on every code PR; adversarial security review when auth or consent boundaries move; a test engineer who distrusts the implementer's summary. Blocking findings loop back — approval binds to a SHA, three rounds max before escalation.
acceptance Preview, not promises
Runs the demo script against the deployed preview. "The backend works" is not acceptance.
release → me Gates, then merge
The release manager verifies approvals, CI, migration safety, and rollback; the merge itself is always mine. Agents have zero production access — read-only included.

Rules learned the hard way

Approvals bind to a SHA. One more push, and the review starts over.
New tests must fail first. A test is only evidence if it failed against the unfixed code — output quoted in the PR.
Three rounds, then escalate. Review loops don't run forever; a human decides what ships with residuals.
Rehearse at production's privilege. A migration that passes as superuser is testing a different system.

Experience

  1. Jun 2026 – now

    Software Engineer Intern · TalkMeUp

    FreshLens computer-vision platform — detection, calibration, evidence evaluation, and the project's closeout gate, shipped under CI, pytest, Ruff, and strict module isolation.

  2. May – Aug 2025

    Software Engineer Intern · Baowin Steel

    Full-stack e-commerce trading platform (Django, React, MySQL) — auth, listings, transactions — deployed to AWS EC2/RDS with CI/CD, migrated off on-prem.

  3. Feb 2024 – Aug 2025

    Undergraduate Researcher · UW–Madison CAT Lab

    React interfaces for a digital-twin traffic monitoring system; D3.js and Chart.js visualizations of live highway data; lazy loading cut map load time ~2s.

  4. May – Jul 2024

    Software Engineer Intern · Lark Technologies

    AI bot automating Google Sheets workflows; multithreading and batch processing turned a 2-week manual analysis of 2,000 customer rows into 4.5 unattended hours.

University of Southern California MS, Computer Science · 2025–2027
University of Wisconsin–Madison BS, Computer Sciences + Mathematics · 2022–2025

Hiring for AI product engineering?

I'm looking for software engineering roles where LLM systems meet production discipline.

0109yizhoulu@gmail.com github.com/RussCTGL LinkedIn