Techbanq

Cloud and platform engineering · Saudi Arabia

Cloud and DevOps consulting in Saudi Arabia

The bill is now unpredictable, nobody has tested the recovery plan, and deployments happen on Thursday evenings because that is when it is safest to break something. We work on the operational half: reliability, recovery, security posture and unit cost.

Structured network cabling running into a switch cabinet

Engagements usually start with a two-week read of the estate — cost concentration, failure domains, recovery position and deployment friction — before anything is changed.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Information security for cloud delivery

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The path to production

What a change passes through before a customer sees it

Six stages. A platform is judged on the last two, and most organisations invest in the first two.

  1. 01

    Source

    Application code and infrastructure definitions in the same review process, because a manual console change is how environments drift apart.

    • Git
    • IaC
  2. 02

    CI/CD

    Build, test, scan and deploy as one gated path, with the same artefact promoted rather than rebuilt per environment.

    • Pipelines
    • Registry
  3. 03

    Infrastructure

    Accounts, networks, identity and data services described in code, so a new environment is a run rather than a project.

    • Terraform
    • Kubernetes
  4. 04

    Application

    Released behind a flag or a progressive rollout, so a bad change is reversed in minutes without a deployment.

    • Flags
    • Rollout
  5. 05

    Observability

    Logs, metrics and traces that answer "is it working for a customer", not only "is the server up".

    • Metrics
    • Tracing
  6. 06

    Reliability and security

    Alerting that reaches a person, an on-call rota that exists, and controls evidenced continuously rather than at audit time.

    • On-call
    • SIEM

Rollback

Reachable from the last three stages and rehearsed rather than documented. A platform where rollback has never been executed does not have one.

Where estates actually hurt

Three symptoms, one underlying cause

The complaints arrive separately — finance asks why the bill grew 40% while traffic did not, the board asks what happens if a region fails, and engineering says a release takes four hours because the rollback is manual. They are usually three views of the same thing: an estate that was assembled to get workloads running and never given an operating model.

That shows up in specific, checkable ways. Resources with no owner tag, so cost cannot be attributed to a team or a product. Environments that were cloned for a test in 2023 and never removed. A recovery plan that exists as a document and has never been executed against production-shaped data. Deployment pipelines that can deploy but cannot reverse, which is why releases cluster on the evenings when the fewest people will notice.

None of it is a cloud provider problem, and none of it is solved by moving provider. It is solved by deciding what the platform guarantees, writing those guarantees as numbers, and building the controls that hold them — which is slower than a migration and considerably more useful.

What we work on

Six areas, in the order they usually matter

01

Architecture and landing zone

Account and subscription structure, network topology, identity boundaries and the guardrails that stop the estate sprawling again. On AWS this is Organizations, SCPs and a landing zone; on Azure it is management groups, policy and hub-spoke. The principle is the same: make the wrong thing hard to do.

02

Migration and modernisation

Workload by workload, with a rollback tested per wave. We are explicit about which workloads are worth re-architecting and which should be moved as they are — rehosting something that will be retired in two years is a reasonable decision, not a failure of ambition.

03

Delivery pipelines

CI/CD that can reverse as easily as it can release: progressive delivery, automated rollback on health signal, and infrastructure defined as code so an environment can be rebuilt rather than repaired. The measure is deployment lead time and change-failure rate, not pipeline count.

04

Kubernetes platform

Where it is justified. Cluster design, namespace and tenancy model, admission policy, resource governance and the upgrade path. Kubernetes solves real problems and creates an operational burden — we say which side of that trade a given estate is on rather than assuming it.

05

Observability and reliability

Service level objectives agreed with the business, error budgets that actually govern release decisions, distributed tracing across service boundaries, and alerts on the business event — a failed payment, a stalled queue — rather than on host metrics nobody acts on.

06

Cost engineering

Allocation first, because you cannot optimise what you cannot attribute. Then waste removal, rightsizing against measured utilisation, storage tiering, data-transfer redesign and a commitment strategy sized to the demand you can actually forecast.

Reliability in practice

What happens when a release goes wrong

The architecture diagram everyone shows is the happy path. This is the one that decides whether Thursday-evening deployments are necessary — and every stage of it is something we build, not something we advise about.

A release flows through four stages: progressive rollout to a small traffic share, an automated health gate comparing error rate, latency and a business signal against the previous version, promotion or halt, and steady state. From any stage before steady state, an automated rollback branch returns traffic to the previous version without a deployment. Database changes are kept backward-compatible so this branch stays available.

  1. 01

    Progressive rollout

    The new version reaches a small share of traffic first, chosen so a fault is visible before it is widespread.

  2. 02

    Automated health gate

    Error rate, latency and a business signal are compared against the previous version for a set period. No human judgement required to spot a regression.

  3. 03

    Promote or halt

    Healthy versions widen automatically. A breach halts the rollout at the current share rather than continuing and alerting someone.

  4. 04

    Steady state

    Full traffic on the new version, previous version kept warm for the rollback window, then retired.

Always available

Automated rollback

At any point before steady state, traffic returns to the previous version without a deploy. This is the branch that has to be tested in a drill, not assumed — an untested rollback is a document, not a control. Database changes are made backward-compatible for exactly this reason: a schema migration that cannot be reversed removes the branch entirely.

Resilience claims

What each claim has to be backed by

Every one of these is routinely asserted in architecture documents and rarely evidenced. The right-hand column is what turns the claim into a fact, and it is what we produce.

"We have a disaster recovery plan"

Design the failover, then rehearse it against production-shaped data with the primary genuinely unavailable.

What proves it: A drill report with the measured RTO and RPO, and the list of things that did not work the first time.

"The system is highly available"

Identify every single point of failure, including the ones outside the application — a shared NAT gateway, one identity provider, a single-AZ database.

What proves it: A failure-domain map and the result of removing each domain in a controlled test.

"Backups are working"

Restore from backup on a schedule and verify the restored data, rather than checking that the backup job reported success.

What proves it: Restore test log with the time taken and the integrity check on the restored dataset.

"The estate is secure"

Continuous posture checks on identity, public exposure, encryption and network paths — plus the segmentation that limits what a compromised workload reaches.

What proves it: Posture findings with severity and owner, and a trend line rather than a single snapshot.

"Costs are under control"

Allocate every resource to a team or product, then report unit cost — cost per transaction, per tenant, per active user.

What proves it: An unallocated-spend figure trending towards zero and a unit-cost series the business recognises.

Platform surface

What we operate with

Grouped by the concern each tool addresses. We are not tied to a single vendor stack, and where a client already runs something workable we improve it rather than replacing it to suit our familiarity.

Cloud platforms

  • AWS
  • Microsoft Azure
  • Google Cloud

Infrastructure as code

  • Terraform
  • AWS CDK
  • Bicep
  • Ansible
  • Helm

Runtime and orchestration

  • Kubernetes
  • Amazon EKS
  • Azure AKS
  • AWS ECS
  • Docker

Delivery

  • GitHub Actions
  • GitLab CI
  • Azure DevOps
  • Argo CD
  • Flux

Observability

  • Prometheus
  • Grafana
  • OpenTelemetry
  • Loki
  • Amazon CloudWatch
  • Azure Monitor

Security and cost

  • AWS Security Hub
  • Microsoft Defender for Cloud
  • Trivy
  • OPA / Gatekeeper
  • Kubecost
  • AWS Cost Explorer

Where to start

What people usually call us about

Each of these has a defined first engagement. None of them requires committing to a platform programme to find out how bad it is.

  1. 01

    The bill grew and nobody can say why

    Usually an allocation problem before it is a waste problem. Two weeks of cost archaeology normally identifies where the growth sits and how much of it is required capacity versus genuinely unused.

  2. 02

    Deployments are frightening

    A symptom of a pipeline that can only move forward. The fix is the rollback branch, backward-compatible schema changes and progressive delivery — normally weeks, not a replatform.

  3. 03

    Nobody has tested the recovery plan

    The most common finding in a Saudi estate review, and the one with the largest gap between the document and reality. A single rehearsed drill converts an assumption into a measured RTO.

  4. 04

    The system is slow and it is not obvious where

    Almost always a data-access or a network-path issue rather than compute. Distributed tracing across the request path usually locates it inside a week, and the answer is frequently a query, not an instance size.

  5. 05

    An audit asked for evidence we do not have

    Access review, change control, encryption state, log retention. Most of it exists in the platform and has never been extracted into a form an auditor accepts.

  6. 06

    One engineer knows how it all works

    A concentration risk rather than a technical one. Infrastructure as code, runbooks and a documented on-call rotation move the knowledge out of one head without a reorganisation.

Common questions

What buyers ask

AWS or Azure — which should we be on?+

For most workloads the decision is made by things other than the platform: existing Microsoft licensing and identity usually favour Azure; a Linux and container-heavy estate with an existing AWS footprint usually favours AWS; data-residency requirements narrow the choice further depending on the region and service. We have not seen a case where the platform itself was the reason a programme succeeded or failed. Multi-cloud as a default position adds real operational cost and is worth it only for a specific, stated reason.

How much can we save on our cloud bill?+

We do not quote a percentage before looking, and we would treat a supplier who does with caution — the honest answer depends entirely on how much of the spend is genuinely unused versus required capacity, and that is not knowable from outside. What we can say is what the review produces: a figure for unallocated spend, a separation of waste from required capacity, and a remediation list ordered by saving against effort and risk. Some estates are already efficient and the finding is that the bill reflects real demand.

Do we need Kubernetes?+

Often not. Kubernetes earns its operational cost when you are running many services with independent release cycles, need fine-grained resource packing, or want portability that is genuinely exercised. For a handful of services with a stable shape, a managed container service or even managed compute is usually cheaper to run and easier to hire for. We would rather tell you that at the architecture stage than operate a cluster you did not need.

Can you work with our existing platform team?+

That is the usual arrangement and the one that leaves the estate in the best state. The pattern that works is that we bring the practices and the initial build, and your team holds the platform from the start rather than receiving it at the end — joint on-call from the first production service, shared runbooks, and infrastructure code in your repository. Platforms handed over at the end tend to be operated nervously.

What about data residency inside the Kingdom?+

Where a workload is subject to a residency requirement, the region choice and the service list are constrained together — not every managed service is available in every region, and a design that assumes otherwise fails late. We establish the applicable requirement and its source with your compliance function first, then design the architecture around what is actually available in-region, including where a service gap forces a different pattern. We do not advise on which regulation applies to your entity; that is your counsel or your regulator, and we build to their answer.

Next step

Review your cloud architecture

A structured review of the estate you already run: where cost concentrates, what happens when the primary region degrades, and how long a deploy takes to reverse. Written findings, whether or not the remediation is ours.