Skip to content
Elena Liu

Hi, I'm Elena,

Trust & Safety Program Manager
who builds the trust layer between humans and AI
with Evals + LLMOps + HITL

At Flip, I used ML classifiers to automate 65% of Tier-1 reports — freeing human experts for the complex, judgment-heavy cases.

This is just the beginning.

Bigger teams. Harder problems. End-to-end ownership.

Ready for what's next.

Work Experience

End-to-end ownership across discovery → specification → delivery → adoption, collaborating closely with AI Research, Engineering, and cross-functional stakeholders.

Trust & Safety Infrastructure
AI/LLM Workflow Automation
Product Lifecycle Management
Data Strategy & Governance
LLMOps & Eval Systems
Cross-functional Stakeholder Alignment

Moody's Analytics

SF Bay Area, CA

Product Operation Program Manager

2024 – Present

Leading Trust & Safety infrastructure modernization — unified 3 legacy moderation tools into a single Internal Safety OS. Partnered with AI Research to deploy an LLM-assisted moderator agent and defined the full automated evaluation strategy.

  • Unified 3 legacy moderation tools into Internal Safety OS — 40% reduction in moderator onboarding time
  • Deployed LLM-assisted moderation agent: 22% accuracy improvement, 15% AHT reduction
  • Defined automated evaluation strategy via Safety Index System tracking Precision, Recall, and FPR
  • Led requirements and prioritization for multi-agent Trust & Safety workflows
  • Extended policy enforcement coverage across advertiser integrity and seller trust domains
Case Study: Moderation OS
40%↓
onboarding time
22%↑
moderation accuracy
15%↓
avg handle time
3→1
tool consolidation

Flip

Los Angeles, CA

Product Operation Specialist

2022 – 2024

Built ML classification pipelines to automate Tier-1 content report processing and supported GenAI policy enforcement tooling.

  • Automated 65% of Tier-1 content reports with ML classifiers
  • Improved moderator decision speed by 12%
  • Supported GenAI content policy enforcement tool pilot launch
  • Managed 100+ backlog items using MoSCoW prioritization
65%Tier-1 automation
12%↑decision speed

LeanData

Sunnyvale, CA

Data Governance Analyst

2020 – 2021

Standardized JSON taxonomy and built Python classification automation, expanding automation coverage 35% and reducing manual reconciliation errors 30%.

  • Standardized JSON taxonomy — unified data definitions across teams
  • Built Python (Scikit-learn) auto-classification system: 35% automation coverage expansion
  • Reduced manual reconciliation errors by 30%
  • Established governance framework for data quality monitoring
35%↑automation coverage
30%↓manual errors

Modis (Adecco / Akkodis)

Remote, CA

Project Manager

2018 – 2020

Managed 4 concurrent IT infrastructure projects with 98% on-time delivery, built real-time Power BI executive dashboards.

  • 4 concurrent IT infrastructure projects — 98% on-time delivery
  • Built real-time executive dashboard in Power BI
  • Cross-functional agile delivery via Jira / Confluence

Projects

yingshill

Moderation OS · Moody's Analytics

Case Study · Production

Turned an LLM moderation assistant from 'it exists' to 'it works.' Unified 3 legacy tools, defined a Safety Index evaluation system — 22% accuracy improvement, 15% AHT reduction, 40% faster onboarding.

LLM ModerationPythonTrust & SafetySafety IndexEval Framework
Read case study →
22%↑
moderation accuracy
15%↓
avg handle time
40%↓
onboarding time
3→1
tools consolidated

ai-content-moderation-edge-case-eval-framework

Open Source

Edge-case evaluation framework for LLM content moderation pipelines — standardized precision/recall tracking, failure pattern taxonomy, reproducible eval suites.

PythonLLM EvalsTrust & SafetyPrecision/Recall
View on GitHub

ai-governance-red-team-control-pipeline

Open Source

Red-team control pipeline for AI governance and policy enforcement — attack pattern classification, prompt injection defense, output consistency test suites.

PythonAI GovernanceRed TeamPolicy Enforcement
View on GitHub

incident-drill-kit

Open Source

Structured incident drill toolkit for T&S ops teams — customizable tabletop exercise scenarios, role assignment templates, and post-incident review frameworks.

Trust & SafetyIncident ManagementOps Tooling
View on GitHub

Education

In Progress · Georgia Institute of Technology

M.S. Computer Science (OMSCS)

In progress — focus on machine learning and computing systems.

2022 · Pepperdine University

M.S. Policy Analytics

Policy analysis and data-driven decision making.

2021 · FullStack Academy

Software Engineering

Full-stack web development.

2018 · Sun Yat-sen University

B.E. Communication Engineering

Communication engineering.

Skills

Languages

Mandarin ChineseNative
EnglishProfessional proficiency

Soft Skills

CommunicationCross-functional LeadershipSystems ThinkingE2E OwnershipBias for ActionInfluence w/o AuthorityDealing with Ambiguity

Tech Stack

AI / LLM
Claude (Anthropic)OpenAIMeta AI (LLaMA)Databricks MLflowGuardrail AI
Safety & Compliance
GDPR / CCPAPII SecurityApache RangerUnity Catalog
Data & Analytics
SQLPython (Pandas / Scikit-learn)ScalaTableauPower BI
Data Engineering
dbtAirflowSnowflakeDatabricksAWS GlueS3
Databases
PostgreSQLMySQLMongoDBCassandra
LLMOps
Eval PipelinesObservabilityLangfuseStreamlit
Platform & Tools
JiraConfluenceGCPTerraformShell ScriptingGitVercel

Personal Projects

yingshill

signal-to-asset

Open Source

Agentic content pipeline that turns a Notion source entry into a linked marketing project, 5 channel-ready drafts, and publish to-dos. Handles source fetch, signal extraction, approval gating, and Notion writeback end to end.

Content OpsWorkflow AutomationKnowledge Ops
PythonClaude CodeNotion APICLIPytest
Artifacts 👇
View on GitHub

course-digest-workflow

Beta

Notion-native course learning system with a modular schema, notebook template, and companion CLI automation that turns course URLs or transcripts into module notes, comprehension prompts, and learning scaffolding via Claude.

Learning OpsWorkflow AutomationKnowledge Ops
Node.jsClaude APINotion APICLISchema Design
Artifacts 👇
View on GitHub

podcast-pipeline

Open Source

CLI pipeline that turns a podcast URL into a structured Notion knowledge base — phrase extraction, cross-episode synthesis, and human eval tracking. Auto-populates Speaker Phrase Library and Podcast & Video Digest databases.

Workflow AutomationAI BuilderKnowledge Ops
PythonGoogle Docs APINotion APILLMHuman-in-the-loop
Artifacts 👇
View on GitHub

design-lore

Open Source

CLI-first design intelligence vault — automated screenshot capture via Puppeteer, CSS/color/typography extraction, built-in trend consultation and personalized challenge generation. Exports design DNA directly to CSS variables.

Product DesignEnablementDeveloper Tooling
TypeScriptNode.jsPuppeteerCLI
Artifacts 👇
View on GitHub

ytm-music-platform

In Build

Personal YouTube Music platform — downloads and stores your library locally via yt-dlp, manages track metadata in SQLite, and plays through a Wavesurfer.js waveform player. FastAPI backend + React frontend for full ownership of your music collection.

Product DesignWorkflow AutomationDeveloper Tooling
PythonFastAPIReactyt-dlpWavesurfer.jsSQLite
Artifacts 👇
View on GitHub

Let's talk

Open to roles in Trust & Safety, AI Program Management, and AI Governance. Based in the SF Bay Area — no sponsorship needed.

© 2026 Elena Liu|Privacy