Skip to main content

Cloud Cost Optimization &
Architecture

We design cloud platforms that scale when you need them, go quiet when you don’t, and show you where every dollar goes.

Cloud architecture and cost optimization services for AWS, Google Cloud and Azure. We design the platform behind your product or your data, from serverless jobs and containers to warehouses and AI services, build it as code in your own accounts, and tune what you already run: right-sizing, scale to zero, storage tiers, commitments and query costs, with a reliability check on every change.

Tran Tien Van, founder of Van Data Team

Tran Tien Van

Founder, cloud and data architect

150+

Projects

Delivered for clients worldwide

3

Clouds

AWS, Google Cloud and Azure in production

5.0/5.0

Rating

Average client rating

71

Upwork Reviews

Client reviews on Upwork

Reference architectures on AWS, Google Cloud, Azure and on-premisesInfrastructure as code in your own accountsSavings with a reliability check on every change

Last updated:

What do cloud architecture services deliver?

Six questions teams ask before they hire someone to design or tune their cloud, answered from the platforms we run.

01 / 06

What does a cloud architecture and cost engagement cover?

The design of the platform and the cost of running it, treated as one job. Depending on where you start, that means a new architecture built as code, or a review and tuning of what you already run.

  • Architecture: compute, networking, data stores, identity, delivery pipelines and observability, drawn and decided before anything is built.
  • Cost: spend mapped to workloads and teams, right-sizing, scale to zero, storage lifecycle, query tuning and commitments sized to real use.
  • Guardrails: tags, budgets, anomaly alerts and a monthly review, so the bill does not drift back.

Key takeaway: Most savings come from how a system is built, not from discounts on what it already wastes.

02 / 06

AWS, Google Cloud or Azure: which should you build on?

Usually the one your team and your data already live in. Moving clouds for a discount rarely pays back once the migration and retraining are counted.

For new builds we weigh the managed services you need (BigQuery is hard to beat for analytics, AWS has the broadest catalogue, Azure fits teams already on Microsoft identity and tools), your existing contracts and credits, and where your data has to stay. Equivalent services are close, not identical: a Google Cloud VPC is global, for example, while AWS and Azure networks are regional.

  • Smaller products often need none of the three: Firebase, Supabase, Vercel or DigitalOcean can run them for less, and we use them when they fit.

Key takeaway: Choose the cloud for the team and the data, then design within it.

03 / 06

How do you review an architecture?

Against the providers’ own frameworks, then against your business. AWS, Google Cloud and Azure each publish a Well-Architected Framework that covers operational excellence, security, reliability, performance, cost optimization and, for AWS and Google Cloud, sustainability.

We go through each pillar with your diagrams, your billing data and your incident history, and rank every finding by risk, effort and saving, so the review ends in a plan rather than a score.

Key takeaway: A review is only useful if it ends in a ranked list of changes you can make.

Source: AWS Well-Architected Framework(opens in a new tab)

04 / 06

Where does cloud waste come from?

  • Managing cloud spend is hard everywhere: in Flexera’s 2025 State of the Cloud survey of more than 750 professionals, 84% named it the top cloud challenge for organizations today.
  • Idle and oversized resources: servers sized for a peak that comes twice a year, dev and staging running all night.
  • Forgotten storage: unattached disks, old snapshots, logs kept forever.
  • Expensive data paths: full table scans on every dashboard refresh, traffic through NAT gateways and across zones.
  • The wrong price: on-demand rates for workloads that have been steady for a year.

Key takeaway: Waste hides in defaults: nobody chose to keep those logs forever.

Source: Flexera, 2025 State of the Cloud Report(opens in a new tab)

05 / 06

How do you cut cost without hurting reliability?

Every change then ships in a small batch, as code, with a reliability check and a rollback, and its saving is measured against the baseline.

  • By naming what we won’t cut before touching anything: capacity that absorbs traffic peaks, compute behind customer-facing latency, databases under live load, backups and failover.

Key takeaway: A saving that causes an incident is not a saving.

06 / 06

How is it priced, and how soon do savings show?

A cost and architecture audit is quoted at a fixed price after a free consultation, sized to your footprint. Implementation is quoted as a fixed bid from the audit, and ongoing reviews run on a fixed monthly scope. We never bill a percentage of savings.

Quick wins, such as idle resources, log retention and right-sizing, usually land in the first weeks; architectural changes and commitments follow over the next two to three months.

Key takeaway: Price and scope are fixed after the audit, before any change is made.

How do we deliver it? Through the four stages below, led by the founder from the first call.

Book a consultation

Reference architectures

Production architectures on every major cloud, and on-premises

Each diagram is the layout we start from for a production platform at a mid-size or large company, in the provider’s own conventions: networks, zones, subnets, security boundaries and the services inside them. Play the flow step by step or switch to the cost view; the note under each names the delivered work closest to it.

Reference architecture · AWS

A multi-AZ production platform on AWS

Edge security in front, a load balancer across two Availability Zones, containers and data in private subnets behind security groups, and identity, keys, audit and backups run centrally.

  1. Route 53 answers for the domain and CloudFront serves the app from edge locations; AWS WAF filters every request with managed rules and rate limits, and Shield Standard absorbs network floods before anything reaches the Region.

    • Users
    • Route 53
    • CloudFront
    • AWS WAF
  2. Customers sign in through Amazon Cognito, which issues short-lived tokens with the tenant in them; the API checks the token on every request instead of keeping sessions.

    • Users
    • Cognito
  3. Requests enter the VPC through the internet gateway and reach an Application Load Balancer in the public subnets of both zones, which ends TLS with an ACM certificate. Its security group accepts HTTPS from CloudFront only.

    • Internet gateway
    • Load balancer
  4. ECS Fargate runs the API in private subnets in both zones and adds tasks as load rises. Each task’s security group accepts traffic from the load balancer alone, Secrets Manager injects credentials at start-up, and outbound calls leave through the NAT gateway of their zone.

    • Fargate tasks
    • Fargate tasks
    • Secrets Manager
    • NAT gateway
    • NAT gateway
  5. Aurora PostgreSQL writes in one zone and keeps a reader in the other, which serves reads and takes over if the writer’s zone fails. ElastiCache holds hot reads; the data tier’s security groups accept connections from the API tasks only.

    • Aurora writer
    • ElastiCache
    • Aurora reader
  6. Files go to S3. Reports and imports go to an SQS queue that Lambda drains, EventBridge runs the scheduled jobs, and the results land back in Aurora, so the API stays fast.

    • Amazon S3
    • Amazon SQS
    • AWS Lambda
    • EventBridge
  7. GitHub Actions assumes an IAM role through OIDC, pushes the scanned image to ECR and rolls the service out; Terraform owns every resource. Engineers sign in with IAM Identity Center and reach containers and the database through Session Manager: no bastion host, no open SSH port, every session logged.

    • GitHub Actions
    • Engineers
    • IAM
    • Amazon ECR
    • Session Manager
  8. Logs, metrics and alarms land in CloudWatch. CloudTrail records every API call and GuardDuty reads it for threats; AWS Backup copies database snapshots to a second Region, and KMS keys encrypt the data at rest.

    • CloudWatch
    • CloudTrail
    • GuardDuty
    • AWS Backup
    • AWS KMS

Closest delivered work: alice/ai runs its API on ECS Fargate with Cognito sign-in, and a healthcare analytics platform ships with Terraform, Secrets Manager and OIDC deploys. Read the alice/ai case study

How we design and tune your cloud

Four stages, each ending in something you can check: a ranked list of findings, a signed-off design, measured savings, and guardrails your team owns.

  1. Stage 01

    Assessment

    Weeks 1 to 2

    We read the architecture and the bill together, with read-only access to your accounts.

    • Billing export: spend mapped to workloads, environments and teams
    • Architecture review against the provider’s Well-Architected Framework
    • Quick wins: idle resources, oversized instances, logs kept forever

    Output: a ranked list of findings, each with its saving, risk and effort.

  2. Stage 02

    Target Design

    Weeks 2 to 3

    A target architecture and the order to get there, signed off before any change starts.

    • Target diagram and a decision log for every service choice
    • What we won’t cut: customer-facing capacity, databases under load, recovery
    • A commitment plan sized to the measured baseline, never to a forecast
  3. Stage 03

    Implementation

    Weeks 3 to 10

    Changes ship in small batches, as code, each with a reliability check and a rollback.

    • Terraform for every change, reviewed in your repository
    • Right-sizing, scale to zero, storage lifecycle and query tuning
    • Savings Plans, reservations or committed use discounts once the baseline is stable

    Output: savings measured against the baseline, change by change.

  4. Stage 04

    Guardrails and Handoff

    Controls that keep the bill from drifting back, then a handoff to your team.

    • Cost allocation tags or labels on every resource
    • Budgets and anomaly alerts per workload, routed to the owner
    • A 30-minute monthly review template and runbooks
    • 30 days of support after handoff

A cloud on autopilot vs an architected cloud

Most cloud bills are not the result of a decision. They grow from defaults, from resources nobody switched off and from prices nobody revisited. This is the difference we build in.

Scaling
Cloud on autopilotServers sized for the peak, all year round
Architected cloudAutoscaling and scale to zero, matched to real traffic
Idle environments
Cloud on autopilotDev and staging run nights and weekends
Architected cloudSchedules or on-demand environments that shut down
Changes
Cloud on autopilotMade by hand in the console
Architected cloudTerraform and CI/CD: every change reviewed and repeatable
Access
Cloud on autopilotLong-lived keys and broad permissions
Architected cloudOIDC deploys, least-privilege roles, a secrets manager
Observability
Cloud on autopilotLogs scattered and kept forever
Architected cloudMetrics, alerts and log retention set per workload
Cost visibility
Cloud on autopilotOne invoice line per service
Architected cloudSpend tagged to workloads, teams and environments
Pricing
Cloud on autopilotOn-demand rates for steady workloads
Architected cloudCommitments sized to the measured baseline
Storage
Cloud on autopilotEvery file in the most expensive tier
Architected cloudLifecycle rules down to cool and archive tiers
Queries
Cloud on autopilotFull scans on every dashboard refresh
Architected cloudPartitioned, clustered tables and cached results
Recovery
Cloud on autopilotBackups nobody has ever restored
Architected cloudTested restores and documented recovery steps

Cloud cost optimization

Cost optimization that follows the architecture

Cloud cost optimization services for AWS, Google Cloud and Azure: we find where the bill really comes from, cut what is waste and protect what keeps you running. These are the six levers we pull, in each provider’s own terms.

The six levers

  • 01

    Right-size compute

    Match CPU and memory to measured use, starting with the largest and idlest resources.

    AWS
    Compute Optimizer
    Google Cloud
    Recommender
    Azure
    Azure Advisor
  • 02

    Scale to zero

    Anything that can wait for a cold start stops costing money between requests and runs.

    AWS
    Lambda, Fargate scaling
    Google Cloud
    Cloud Run services and jobs
    Azure
    Container Apps, Functions
  • 03

    Commit to the baseline

    Commitments cover only the usage measured as steady, never a forecast.

    AWS
    Savings Plans, Reserved Instances
    Google Cloud
    Committed use discounts
    Azure
    Reservations, savings plans
  • 04

    Tier the storage

    Lifecycle rules move data to cheaper classes as it ages and delete what nobody needs.

    AWS
    S3 lifecycle, Glacier
    Google Cloud
    Nearline, Coldline, Archive
    Azure
    Cool and Archive tiers
  • 05

    Tune the warehouse

    Partitioned, clustered tables and columnar files make every query read less data.

    AWS
    Athena with Parquet, Redshift
    Google Cloud
    BigQuery partitions, clustering
    Azure
    Databricks job clusters
  • 06

    Watch the network

    Data transfer hides in NAT gateways, cross-zone traffic and egress; endpoints and placement cut it.

    AWS
    NAT gateway, VPC endpoints
    Google Cloud
    Cloud NAT, egress
    Azure
    NAT Gateway, Private Link

A loop, not a one-off sprint

  1. 1Inform. See the spend: costs tagged to workloads, teams and environments, against a baseline.
  2. 2Optimize. Act on it: right-size, schedule, tier storage and tune queries, then commit to the steady baseline.
  3. 3Operate. Keep it: budgets, anomaly alerts, a monthly review and guardrails on every new workload.

Source: FinOps Foundation: FinOps phases(opens in a new tab)

What we won’t cut

  • Capacity that absorbs your traffic peaks
  • Compute behind customer-facing latency
  • Databases under live query load
  • Backups, disaster recovery and multi-zone failover
  • Anything that would make your team tiptoe around the platform later

We never bill a share of the savings: a percentage rewards cutting too deep.

Book a consultation
Tran Tien Van, founder of Van Data Team

Warning signs

Signs your cloud needs an architect, not another discount

If two or more of these sound familiar, the bill is telling you something about the architecture.

  • 01The cloud bill grows faster than your users or your revenue.
  • 02Nobody can say which team or feature a cost line belongs to.
  • 03Dev and staging run around the clock at production size.
  • 04A forgotten resource, snapshot or log group turns up on the invoice months later.
  • 05Infrastructure is changed by hand in the console, and nobody can rebuild it.
  • 06Your warehouse or AI bill jumps whenever someone ships a new dashboard or prompt.

If you're seeing any of these, it's time to talk.

Book a consultation

Cloud work

Platforms we built and run, on every major cloud

Named products where the clients have agreed, and unnamed engagements otherwise. Every number comes from the project itself.

Production products we architected and shipped
20+
Lower LLM cost after one model change
~92%
Clouds with systems we built in production
3
youheal.com
YouHeal home page
YouHeal on a phone
YouHeal logoClient projectLive

Precision medicine

YouHeal Precision-Medicine Platform on Azure

An Azure-native platform with a client portal, practitioner tools and a personalised AI health assistant: rule-based engines produce the analysis, and the AI answers from each client's own reports and the clinic's knowledge base.

  • Client portal, practitioner tools and an AI health assistant on Azure
  • AI answers grounded in each client's own reports, under practitioner control
  • Analysis from rule-based engines; the AI explains results but cannot change them

Azure · Azure AI Services · Azure AI Search · FastAPI · React · TypeScript · PostgreSQL · Terraform · GitHub Actions

More cloud engagements

Focused work on cost and architecture. The clients stay unnamed; the services are what shipped.

  • Sales KPI pipeline on Cloud Run

    About 92% lower LLM cost after moving extraction to Claude Haiku, on Cloud Run Jobs that bill only while they run.

    Cloud Run Jobs · BigQuery · Claude Haiku

  • EC2 to Lambda scraper migration

    A headless Chromium scraper moved from always-on EC2 to a Lambda container that scales to zero between runs.

    AWS Lambda · ECR · Docker

  • Multi-tenant Shopify hub

    Four single-purpose Cloud Run services; webhook-driven sync replaced the daily full re-pull and its compute.

    Cloud Run · BigQuery · Firestore · Pub/Sub

  • Healthcare ETL on Glue

    Glue capacity scales with the number of sources inside set bounds, and Parquet keeps Athena scans small.

    Airflow · AWS Glue · Athena · S3

  • Healthcare analytics on Fargate

    Terraform-managed ECS Fargate with Secrets Manager and OIDC deploys, and 255+ tests before anything ships.

    ECS Fargate · Terraform · Secrets Manager

  • Common Crawl extraction on EMR

    PySpark on EMR reads Common Crawl’s public archives instead of running a crawler, in resumable batches.

    AWS EMR · PySpark · S3

  • Lambda integration suite

    Four SaaS integrations moved off always-on servers into stateless Lambda functions with SNS failure alerts.

    AWS Lambda · RDS MySQL · SNS

  • Five campaigns, one Cloud Run service

    A single dispatcher service replaced five scheduled deployments for five revenue-reporting campaigns.

    Cloud Run · Cloud Scheduler · BigQuery

What clients say about our cloud and data work

Teams bring us AI agents, web scraping, data platforms, reporting and cloud work. Their verified reviews speak to senior, founder-led delivery from prototype to production.

Verified on Upwork

5.0

71 client reviews · 69 of them five-star

Top Rated Plus · 100% Job Success · 100+ jobs

  • 5 stars69
  • 4 stars2
  • 3 stars0
  • 2 stars0
  • 1 star0
See Upwork Reviews
CK

Chris K.

December 30, 2025

“Great backend developer, highly recommend!”

FT Platform Phase #2

Verified Upwork review

GB

Gilad B.

December 18, 2025

“Very knowledgeable and professional. Good communication”

Phase 0: Design a granular data schema and structure, and full tool flow

Verified Upwork review

JD

Julio D.

July 16, 2025

“Tran was great, very knowledgeable and quick responses”

Web scraper in R

Verified Upwork review

MM

Madison M.

March 31, 2025

“Very responsive and quick to get started! Produced excellent results. I will definitely reach out again in the future.”

Review Git Pull Requests

Verified Upwork review

DC

David C.

March 29, 2025

“Great platform to interface with developer.”

You will get AWS, GCP and Azure Data pipeline

Verified Upwork review

AJ

Alex J.

February 18, 2025

“Fast, responsive, professional. Really appreciated the thorough documentation too.”

You will get Data Scraping | Data Extraction | Web Scraper | Automation Tools

Verified Upwork review

TS

Tomer S.

December 12, 2024

“Van exceeded all expectations with exceptional professionalism and expertise. They delivered high-quality work ahead of schedule, communicated effectively throughout the project, and made the collaboration seamless and enjoyable. I highly recommend Van to anyone looking for a skilled and reliable freelancer.”

Create Web scraper for Facebook

Verified Upwork review

PT

Preska T.

October 21, 2024

“I recently had the pleasure of working with Tran, and I can't express enough how impressed I am with his work. From the very beginning, he demonstrated a deep understanding of our project requirements and brought a level of expertise that made a significant difference in the outcome. What truly sets Tran apart is his commitment to excellence.”

You will get Data Scraping | Data Extraction | Web Scraper | Automation Tools

Verified Upwork review

AW

Alice W.

May 26, 2024

“AMAZING. We are lucky that we found Van. He helped us with our database structure. He is very knowledgeable and very cooperative. We are still continuing to work with him further. I can only highly recommend.”

Admin Panel for Data Management

Verified Upwork review

NS

Nic S.

May 15, 2024

“Excellent work. I'd be very happy to work with Tran in the future.”

Web scraping: Project review and proposal

Verified Upwork review

PK

Paris K.

March 22, 2024

“Tran did a great job on a LAMP REST API deployment to Microsoft Azure. We'd be happy to work with this freelancer again.”

Seeking developers experienced with LAMP (Python) and REST APIs for UX Research Study / gstd-2024-1

Verified Upwork review

YK

Yalcin K.

March 18, 2024

“Van is a great data engineer, and I highly recommend it. He joined our project and helped build a custom data pipeline within weeks.”

Python Developer to Build a Shopify Integration

Verified Upwork review

The Van Data Team difference

Why teams trust us with their cloud

Five reasons the work holds up after we leave.

Tran Tien Van, founder of Van Data Team

Tran Tien Van

Founder, AI and software architect

Top Rated Plus on Upwork
The engineer who scopes your project is the one who builds it.
Projects delivered
150+
Upwork rating from 71 reviews
5.0
Countries served
15+
Years of development experience
5+
  1. 01

    Architects who write the Terraform

    The engineer who reviews your architecture is the one who changes it, so decisions move at the speed of a pull request.

  2. 02

    Savings that respect reliability

    Every change comes with a reliability check and a rollback, and we name what we won’t cut before touching anything.

  3. 03

    No share-of-savings billing

    Fixed bids and retainers only. A percentage of savings rewards cutting too deep; we would rather be paid to get it right.

  4. 04

    Three clouds, one method

    Systems in production on AWS, Google Cloud and Azure, so recommendations come from shipped platforms, not slides.

  5. 05

    Guardrails included

    Tags, budgets, alerts and a monthly review ship with the work, so the bill does not drift back once we are gone.

150+ Projects Delivered With VANDATATEAM

VanDataTeam has delivered 150+ projects for teams across 15+ countries, building lasting success through production grade AI Agent systems and data engineering expertise.

  • athenahealth
  • YouHeal
  • OBJX Intelligence
  • Zeno
  • Yōjin
  • Consequential Flowers
  • Clear Gene
  • Supply Bridge
  • Conversion Finder
  • Envoy
  • Lunada
  • Ahrvo
  • Ellipsis Earth
  • Longpoint
  • Intelligent World
  • Debit My Data
  • Graspr AI
  • SeaSure
  • alice/ai
  • Setmore
  • White Ribbon Alliance
  • HBC
  • Pareto Intelligence
  • Credit Club
  • Kejora Capital
  • iRASSHAi
  • Wajooba
  • Voodoo
  • Forskningslogen Friederich Münter
  • Flight IQ
  • Star Analytics

FROM DELIVERED PROJECTS

The cloud stack we work with

205 technologies across 14 layers, every one of them used in delivered client work.

Pick a layer, or let the orbit walk you through them.

Cloud Platforms

14 tools

The three major clouds, Oracle, Alibaba and IBM, and leaner platforms where they fit better.

  • AWS
  • Google Cloud
  • Azure
  • Oracle Cloud
  • Alibaba Cloud
  • IBM Cloud
  • DigitalOcean
  • Cloudflare
  • Vercel
  • Firebase
  • Supabase
  • Heroku
  • Red Hat OpenShift
  • CapRover

On the orbitAWS · Google Cloud · Azure · Alibaba Cloud · DigitalOcean · Cloudflare · Vercel · Firebase · Supabase

Cloud architecture and cost FAQ

Show us your architecture and your bill

Book a free 30-minute call with the founder. Bring a recent invoice or billing export and a rough diagram, and we will tell you where we would look first and what we would leave alone. No deck, no pressure.

We use these details only to reply about your project. See our privacy policy.

Tran Tien Van, Founder of Van Data Team, ready to help

Keep exploring

Services teams often pair with this one, delivered by the same founder-led team.

See all 6 services