I’m Xiang Zhu, a Cloud and DevOps Engineer. I bring 16 years of enterprise IT
experience, including consulting roles at Thoughtworks and Accenture, to
infrastructure design, automated delivery and systems operations. I also build LLM
workflows and use AI coding agents in everyday engineering, checking changes against
tests and running systems.
I’m looking for a hands-on DevOps, Platform Engineering or SRE role in New
Zealand.
Full work rights (open work visa).
Selected engagements from six years as a consultant at Thoughtworks, across banking,
fintech, automotive, home services and public health. How I got there, from 2006
onwards, is on my career journey page.
Supported Azure/AKS and development teams day to day. Led and completed the move of SIT
and UAT into the government-managed Azure subscription that hosted production, to meet
mandatory security requirements, and rebuilt Bamboo delivery in GitLab CI/CD, guiding
another DevOps engineer.
Led the initial assessment and the platform design in the first two phases for a
bank where VMs and containers had to coexist, and designed predictable IP addressing
for business applications. The platform went on to support 1,000+ business applications. Later,
business analysis for operations systems connected operational needs with monitoring.
Supported AWS/EKS infrastructure for a fintech platform with 20+ microservices and over
a million active users. Provisioned resources with Terraform, including modules
intended to support disaster recovery, and helped teams deploy through the existing
Argo CD and GitHub Actions setup.
Refactored Terraform modules and built service modules for a recovery environment
while supporting an AWS platform. Together with a colleague, I assessed platform risks
and prioritised operational improvements.
Migrated gopass credentials to Vault on OpenShift, reducing password synchronisation
from over an hour to seconds for more than 200 users. Implemented GoCD
Pipeline-as-Code and supported service migrations, platform operations and handover.
Does another review step
produce a better knowledge article?
A master's capstone comparing single-pass prompts and an agent review loop for
technical support records. Model orchestration, recoverable batch runs, evaluation
design and interpretation of results.
Single-pass CoTChange the prompt; keep one context
Simplify
69 selected technical records · Offline research
Saved scores, checked interpretations.
How I work with AI
Build · Delegate · Verify
I work with AI coding agents, mainly Codex and Claude Code, and I have built and evaluated
LLM workflows. I set the scope and constraints, let the agent investigate and implement,
and accept a change only after checking it against source, tests and running behaviour.
In my master’s capstone I built a LangGraph generator–auditor loop on a local Ollama
backend, then used paired scoring across 69 technical records to test what the extra
review step delivered.
AI helped trace dependencies and source findings. I set the boundaries, and the
release went out after 522 automated tests, seven recovery scenarios and fresh scans
passed.
Claude Code reviewed this site and implemented link previews, search files and a
mobile header fix. Browser checks at 320px exposed an overflow, and I spotted a
missing changelog entry before the change went live.
Why useful work can go undone while people able to do it stay out of work: a social
need is not a funded job, and a hiring round that ends without an offer cannot, by
itself, show that talent is missing.
How a Trivy scan stopped a Story Filler release, and how dependency updates, a smaller
runtime image and API smoke checks helped verify the fix without weakening the
security gate.
Reflections from AI research and engineering practice on human judgment, inspectable
outputs and the checks needed to support a decision, with verification as part of the
working process.
A follow-up on separate frontend and backend delivery paths in GitHub Actions, with
Terraform for cloud configuration and remote state. The current case study adds
implementation details and verified release records.
The original migration account follows a personally maintained application from a
Kubernetes lab to Cloud Run, Neon and static hosting, weighing infrastructure control
against the work of operating it alone.
How regularly I used Codex across selected project workspaces. The counts show activity,
not quality; How I work with AI shows how the results were checked.
Recorded Codex work cycles across selected software projects. Snapshot updated 3
November 2026; the final day is partial.
What these numbers include
Aggregated locally from retained Codex metadata for explicitly selected
software-project workspaces. A turn is one recorded assistant work cycle; a
follow-up starts another turn. Planning and documentation in those workspaces are
included. Other workspaces, including job-search tasks, are excluded.
Only top-level turns with a start time and a finished, failed or interrupted status
are counted. Forked copies are deduplicated by turn ID; subagent work is excluded.
Dates use each turn’s start in Pacific/Auckland. This export omitted 0 duplicate
records, 0 records without a start time and 2 nonterminal records. An active day has
at least one qualifying turn; zero means no qualifying record was retained.
I generated this snapshot from retained records in the selected local workspaces. It
reports daily activity in recorded turns; the project pages show the resulting code,
decisions and verification. The public export contains dates and aggregate counts.
Conversation content and workspace details remain local.
My background spans 16 years in enterprise IT (2008–2024), including consulting, systems
engineering and project delivery. My work connects cloud platforms, container
orchestration, delivery automation and day-to-day operations.
At
Thoughtworks (2018–2024), I worked as a Senior Consultant on
infrastructure and DevOps projects for clients including Singapore’s Health Promotion
Board, Daimler AG, China Merchants Bank and SadaPay. Earlier, at
Accenture (2016–2017), I was seconded to Huawei as an Infrastructure
Consultant. I was the only Kubernetes specialist on the HiCloud container platform,
designing the container cluster solution and helping the client team build its
Kubernetes capability.
I have completed a
Master of Software Engineering at Yoobee Colleges
in Auckland. This portfolio complements my enterprise career with recent, inspectable
practice in infrastructure and cloud delivery. I use AI to support investigation and
implementation, then check changes against source, tests and deployed behaviour.