ARCH — Architecture Control Hub
A visual cloud architecture designer that emits reviewable Terraform for AWS, Azure and Google Cloud.
Gokul Upadhyay Guragaingocools
DevOps and MLOps Engineer
Gokul Upadhyay Guragain
I build the delivery paths that get software into production and keep it there. Cloud infrastructure as code, pipelines that fail loudly, and observability that answers questions before anyone thinks to ask them.
Real profile media is served from R2.
At a glance
Positions, interests, and the work between them
Positions
- Trainee DevOps EngineeracAIberryMar 2026 — Apr 2026
- Fellow AI/MLOps EngineerFusemachinesMay 2025 — Dec 2025
- Apprentice Solutions ArchitectAdex InternationalNov 2024 — Feb 2025
Interests
- Cloud architecture
- Infrastructure as code
- Delivery systems
- MLOps
- Observability
- Teaching and community
Grounded in the projects, skills, and community work published here.
Current focus
Building ARCH, an architecture-as-code control hub that turns a drawn cloud topology into reviewable Terraform. Reading about eBPF-based network observability. Preparing the next AWS Student Builder session at LBEF.
Read the latestWhat I build
Three things, repeatedly
The work divides cleanly. Everything else I do is in service of one of these.
Infrastructure that is described, not assembled
Terraform as the primary artefact rather than documentation written afterwards. Composable modules, remote state with locking, and availability decided per failure domain with the cost written down.
Cloud notesDelivery paths that fail loudly
Pipelines that produce evidence as well as artefacts, deployments identified by image digest rather than a mutable tag, and a production gate that can actually say no.
DevOps notesModel serving treated as production
An inference service is a production service with one extra failure mode: it can be healthy, fast and wrong. Pipeline-driven training, metric-gated registration, and monitoring that pairs request metrics with prediction distribution.
MLOps notes
Selected work
gocools learn
A cloud learning platform with structured courses, practice examinations and certification tracks.
gocools write
A writing and documentation workspace with grammar assistance, AI-text detection and document management.
Compliant delivery pipeline for a healthcare platform
GitLab CI pipelines for a HITRUST-regulated healthcare system, where every stage has to produce evidence as well as an artefact.
What I work with
Tooling
Full breakdownGrouped by what the tool is for. Core means I have run it under real load and know how it fails.
Cloud
- AWS
- AWS SageMaker
- Microsoft Azure
- Cloudflare
Containers and orchestration
- Docker
- Kubernetes
- K3s
- Helm
- kind and minikube
Infrastructure as code
- Terraform
- Ansible
CI/CD and GitOps
- GitHub Actions
- GitLab CI
- Argo CD
- Git and GitHub
Experience
Where the work happened
Full recordMar 2026 — Apr 2026
Trainee DevOps EngineeracAIberry
Delivery and MLOps work on a healthcare platform operating under HITRUST, where the pipeline is an audited control rather than convenience tooling.
May 2025 — Dec 2025
Fellow AI/MLOps EngineerFusemachines
Containerised ML pipelines and K3s-hosted inference, instrumented so that a model service is treated as a production service with an extra failure mode.
Nov 2024 — Feb 2025
Apprentice Solutions ArchitectAdex International
AWS architecture work where availability, cost and blast radius were treated as one conversation, and Terraform was the primary artefact rather than an afterthought.
Jun 2024 — Oct 2024
Cloud Engineering TraineeMentor Me Collective
Structured cloud engineering training built around practical tasks rather than lecture material.
Writing
Recent articles
All writingThe pipeline is the control
In a regulated environment a CI pipeline is not automation that happens to be convenient. It is the evidence that software reached production legitimately, and evidence has to outlive a log retention window.
Notebooks may explore, they may not deploy
The line that made model provenance tractable was not banning notebooks. It was deciding that nothing which reaches an endpoint may have originated in one.
A model can be healthy, fast and wrong
Standard service monitoring tells you an inference endpoint is up and responsive. It cannot tell you the predictions stopped making sense three days ago.
Notes
Shorter things
All notesThings I did not want to work out twice.
No apply button in a web UI
A convenience feature that mutates production is a blast-radius decision wearing a convenience costume.
Image digests over tags, always
A tag is a mutable pointer. Deploying by tag means you cannot state what is running.
Terraform state locking is not optional
Two engineers, one state file, no lock. The failure is exactly as bad as the documentation implies.
Accepted findings need a home
If there is no recorded way to accept a scan finding, engineers will find an unrecorded way.
Now
Building ARCH, an architecture-as-code control hub that turns a drawn cloud topology into reviewable Terraform. Reading about eBPF-based network observability. Preparing the next AWS Student Builder session at LBEF.
Contact
If any of this is close to a problem you are trying to solve, I would like to hear about it.