I'm DrOSAlchemist, an independent systems engineer and network engineer. My practice connects DevSecOps, SRE, Kubernetes, model infrastructure and secure network design: the infrastructure, identity, automation and policy that make a platform useful after the demo is over.
I'm pursuing a Doctor of Information Technology (DIT). Alongside that study, my engineering questions focus on latency, failure, identity boundaries, policy enforcement and operational drift. I use the lab to test assumptions and separate what I expect a system to do from what the evidence actually shows.
I like problems that cross boundaries. A model needs a GPU, a scheduler needs a capacity contract, a service needs a name and an address, and a secure network path. Every change needs a way to be observed, validated and reversed. Connecting those details is where reliable systems take shape.
This portfolio connects implementation to evidence: source code, test contracts, dated measurements and explicit limitations. It distinguishes runnable demos from reference architectures, and design goals from verified production behavior.
This is also my engineering lab: a place to turn a systems question into a bounded experiment, test the failure and refusal paths, and make the result inspectable. I ask which contract defines the boundary, which control enforces it and which network conditions affect the outcome. I prefer a small, reproducible claim to a large architecture story with no working verification.
The lab is a collection of source-backed exercises, not a hosted cloud sandbox. Commands execute on the reader's machine; this website does not run code or provision resources.
Frame a falsifiable question
Write the input, invariant and rejection condition. Define what observation would disprove the proposed behavior before changing code.
Bound the environment
Record the revision, runtime, tool versions and fixture assumptions. Start locally; credentials, billable infrastructure and destructive changes need a separate decision.
Exercise the failure path
Run the existing tests, inspect refused inputs and compare the expected result with the actual exit code. A green happy path alone is not enough.
Publish evidence and uncertainty
Keep commands, logs and findings together. Redact secrets, state what was not tested, and distinguish an observed result from a hypothesis or next experiment.
Workbench / reproducible starting points
Pick a question. Inspect the result.
Use a new directory for each clone and inspect the source before running commands. Start with local verification or read-only research. Expand an exercise only after reviewing its dependencies and boundaries.
Can declared infrastructure and runtime contents be checked before obtaining cloud credentials?
Prerequisites
Git, Node.js 24 or newer and npm. Internet access is needed for the initial clone. The extended infrastructure track also requires Terraform 1.16.x and the pinned OPA installer.
Execution scope
The starter below runs application and plan-extractor tests only. It neither installs scanners nor runs Terraform, builds an image or contacts Azure.
Reproduce / azure-devsecops-policy-pipeline
git clone https://github.com/DrOSAlchemist/azure-devsecops-policy-pipeline.git
cd azure-devsecops-policy-pipeline
git checkout --detach 4aefdf1de98fa4742ab42dee15c388f6e59cb90e
npm ci --ignore-scripts
npm run check && npm test
Expected observation
At this pinned revision, the Node suite reports 11 passing tests. Inspect route behavior and extractor refusal cases in the test output; this command does not reproduce the full CI scan.
Failure check
Inspect tests/extract-plan.test.js for malformed events, duplicate plans and pre-existing output. The extractor must not overwrite old evidence or accept an empty plan.
Next experiment
Follow the implementation walkthrough for the separate Terraform mock-plan, OPA and scanner tracks. Preserve tool versions and capture the generated plan, denial reasons and exit codes.
Boundary
Mock plans cannot prove Azure RBAC, actual Policy enforcement or application deployment. The template's default base image is not the built application. Do not run terraform apply as a lab step.
Does the site distinguish an unavailable repository feed from a genuinely empty account?
Prerequisites
Git, Node.js 22 or newer and npm. Dependency installation needs internet access. Use a fresh directory; the rendered checks start and stop their own loopback server.
Execution scope
Tests run with synthetic GitHub responses in an isolated subprocess. No GitHub token, hosting account, production outage or cloud resource is required.
Reproduce / dr-os-alchemist-portfolio
git clone https://github.com/DrOSAlchemist/dr-os-alchemist-portfolio.git
cd dr-os-alchemist-portfolio
git rev-parse HEAD
npm ci
npm test && npm run lint && npm run build && npm run test:rendered
Expected observation
Every command must exit successfully. The rendered checks verify curated evidence during feed failure, repository controls during success, and static article routes. Record the printed revision because main changes.
Failure check
The feed suite injects timeouts, rate limits, malformed data and a later-page failure. Failed lookups must return unavailable with no-store, not a partial count or a successful empty list.
Next experiment
In a separate terminal, run npm start after the build, open localhost:3000, then exercise combined filters, keyboard skip navigation and narrow-screen layouts. Stop the server when finished.
Boundary
Fixtures do not verify GitHub uptime, Netlify cache behavior or external-link availability. Local checks do not establish a security audit or production service-level objective.
What did the reported 659-to-338-second result measure, and which variables changed together?
Prerequisites
Git and internet access to read the public research record. No GPU or cloud account is needed for this inspection.
Execution scope
The commands download source and print the research note; they do not install dependencies, provision a cluster, start inference or reproduce the benchmark.
Reproduce / keda-scale-zero-gpu-inference
git clone https://github.com/DrOSAlchemist/keda-scale-zero-gpu-inference.git
cd keda-scale-zero-gpu-inference
git rev-parse HEAD
git show HEAD:docs/cold-start-optimization.md
Expected observation
Identify the dated 2026-04-05 T4 Spot record in us-east1-d, the queue-spike-to-first-token boundary, the baseline and the combined storage/boot-disk change.
Failure check
Check whether an interpretation incorrectly attributes the improvement to persistent storage alone or generalizes it to L4 GPUs. Neither conclusion is supported by the cited measurement.
Next experiment
Design a controlled follow-up with fixed model, image, GPU, region and cache state, repeated trials and separate provisioning/load/inference timings before buying capacity.
Boundary
Persistent model storage and a Secondary Boot Disk were measured together. This is source inspection, not a rerun; any cloud reproduction needs separate approval, budget limits and cleanup.
Question, source revision, date, environment and exact commands.
Expected versus observed behavior, exit codes and refusal cases.
Raw evidence location, retention limits and changed variables.
Known gaps, cleanup performed and the next falsifiable question.
Do not collapse the evidence levels
A local test verifies its assertions under the tested inputs.
A CI run records one revision and one execution environment.
A cloud measurement describes its dated setup, not every workload.
A design proposal remains a proposal until implemented and verified.
Never publish tokens, private logs or customer data as lab evidence. No exercise here claims production availability, compliance certification or a complete security assessment.
DR OS / SYSTEMS PRACTICEProfile image via GitHub
01
Make the system legible
Good automation exposes intent, ownership and failure modes instead of hiding them behind a button.
02
Design for the day after launch
Operations, observability and rollback are part of the architecture, not cleanup for later.
03
Secure the seams
Identity, policy and network boundaries matter most where components meet.