← back
Skills
- Languages
- TypeScript, JavaScript, Python, Bash, SQL
- AI & Agents
- Agent orchestration, LLM evaluation, tool calling, context engineering,
system-prompt design, structured output, MCP, Claude Code
- Web & Backend
- React, Next.js, Node.js, REST APIs, PostgreSQL
- Testing
- Playwright, Appium, Page Object Model, CI test infrastructure, fault injection
- Infrastructure
- Docker, GitHub Actions, CI/CD, Git, Linux/WSL2
Work Experience
QA Wolf
QA Lead
10/2025 to Present
- Lead QA for Harvey, an $11B AI legal platform: own test strategy and the client
relationship, scope and outline new coverage with their CSMs and product stakeholders in
recurring client meetings, and own the delivery reporting they see
- Maintain 1,600+ Playwright/TypeScript E2E tests across 13 product domains on an
82-class Page Object framework (~76K lines) with a compiler-enforced PageFactory registry;
executed platform-wide POM codemod and environment migrations across the full suite
- Made failure evidence machine-consumable for an AI agent: run logs, Playwright traces,
failure video and live application DOM pulled programmatically instead of screenshotted by a human,
the integration that made autonomous triage possible
- Triaged failures by root cause rather than one test at a time: a single call-graph census pinned a
suite-wide defect across 71 affected tests in one pass, and a re-run-first policy proved a third of
batch failures were infrastructure contention needing no code change
- Converted ~1,300 client spec cases into a machine-readable ground-truth dataset, then gated
every test against its own spec steps before shipping, so coverage was verified rather than assumed
QA Wolf
QA Engineer
07/2025 to 10/2025
- Promoted to QA Lead within three months of hire on delivery metrics
- Validated non-deterministic LLM behavior across 65 tests: agentic document workflows, streaming
completions with multi-minute retry windows, citation integrity and knowledge-source attachment, spanning
native PC, Word Web, iOS/Appium, email and Twilio SMS/MFA, and stealth-browser OAuth against anti-bot protections
Zeta Global
Software Developer
02/2023 to 07/2025
- Built internal React/Next.js dashboards with real-time PostgreSQL filtering, adopted
across teams as their working view into the data and retiring a standing manual reporting job
- Built a Python collection service over REST APIs and SQL that consolidated scattered
sources into one validated pipeline, replacing manual gathering and measurably improving the
accuracy of the data teams reported on
- Owned the deployment path end to end: Git, GitHub Actions and Docker pipelines that
tested and deployed both systems
Projects
The Control Room
human-in-the-loop AI automation
Personal project, 2026 to Present
kxvin1.github.io/control-room
- Cut a broken test from ten manual hand-offs down to two: an AI agent owns the repair, four
always-on panels surface only the steps that still need a person, and the day's reading load drops from
thirty failures to two decisions
- Designed evidence gates so the agent cannot self-certify: "ready" requires two clean local runs
plus one infrastructure run recorded against the same commit, run id attached for audit.
135 automated self-checks guard the harness itself, including fault-injection tests that must catch
a planted failure before a checker is trusted
- Cut agent context cost from 3-8k tokens to ~1k per investigation by replacing raw page snapshots
with filtered evidence reports, and live-DOM inspection from 62s to 17s, so long investigations stopped
exhausting context; the same collected evidence turns a client-facing bug report from 20 to 30 minutes of
writing into one keystroke
- Architected panels and agent as decoupled processes communicating only through plain-text files,
no server and no API, so either side restarts mid-round without losing state; a sentinel process exits the
instant a tripwire fires, surfacing unattended failures in ~15s rather than at the next check-in, and a
supervisor verifies panel relaunches and reports crash loops as bugs instead of retrying forever
- Built a self-improvement loop: the agent files its own bugs from work it encounters, runs a nightly
retro scoring past predictions against outcomes to calibrate confidence, and converts each lesson into a
version-controlled rule its next session starts with, every entry stamped with human or AI authorship and
enforced by a test
Education
App Academy1000-hour full stack web development, 07/2022
·CSU NorthridgeB.S. Psychology, 12/2019