How Diaflow Built Active-Active AWS and SOC 2 Readiness in 4 Weeks

Diaflow needed three-region AWS infrastructure and SOC 2, HIPAA and GDPR readiness at the same time to close enterprise deals. Running architecture and compliance in parallel against one Terraform codebase got both done in 4 weeks, with manual operations work down 80% and 99.9% uptime. The post walks a CloudFront and S3 Content-Type incident end to end.

Customer storyUpdated Nov 2, 2026

Cover Image for How Diaflow Built Active-Active AWS and SOC 2 Readiness in 4 Weeks

How Diaflow Built Active-Active AWS and SOC 2 Readiness in 4 Weeks

Every enterprise security questionnaire asks two things early: show us your SOC 2 report, and tell us how you fail over between regions. For a fast-growing SaaS platform, those two answers are the gate between today's customers and the next tier of contracts.

Diaflow needed both at once, and the usual route does them in sequence: build the multi-region architecture, then audit it, then reopen whatever the audit finds. With CloudThinker, Diaflow ran architecture and compliance in parallel against one Terraform codebase and stood up three-region infrastructure with SOC 2, HIPAA and GDPR readiness in 4 weeks. Day 2 then went to CloudThinker agents, with Resolve investigating incidents across the stack.

The customer

Diaflow is an AI-native automation platform on AWS. To close enterprise deals, three requirements landed at the same time:

  1. Global latency. An active-active deployment across three AWS regions (US, EU and APAC), with no downtime during the move.
  2. Regulatory readiness. SOC 2, HIPAA and GDPR controls in place to unblock the enterprise pipeline.
  3. Velocity. None of it could slow down feature shipping.

Why the traditional route takes months

The usual answer is to hire specialized cloud architects and security engineers, and the hiring alone takes longer than the market will wait. The work itself is also structured to be slow:

Architecture and compliance run in sequence. A team designs the multi-region topology, builds it, and then a compliance effort audits what was built. Every finding (encryption at rest, IAM least privilege, logging) means reopening infrastructure that's already live.

State is the hard part, and it comes late. Moving stateless services between regions is routine. Replicating databases with acceptable lag and clean failover is where projects stall, and it's usually discovered after the easy wins are done.

Compliance evidence is a spreadsheet exercise. Controls get documented once for the audit and drift the week after. The next audit starts from scratch.

Day 2 isn't planned. A multi-region footprint doubles the surface to monitor and the bill to watch. Without automation, that load lands on the developers who are supposed to be shipping features.

Where CloudThinker fits

Diaflow worked with CloudThinker instead of adding contractors. The work ran as four workstreams against the same architecture, so compliance shaped the design instead of auditing it afterwards.

01Migration

Active-active

Cloud-native topology across three regions, moved without downtime.

02Optimization

Right-sized

Right-sizing, resource scheduling and Savings Plan recommendations for the new footprint.

03Security

Continuous

Vulnerability scanning and misconfiguration detection with automated remediation.

04Compliance

SOC 2, HIPAA, GDPR

Monitoring, evidence and policy enforcement running in the background.

Four workstreams run in parallel against one architecture, so compliance shaped the design instead of auditing it afterwards.

The phases overlapped on purpose, so each workstream could inform the others instead of producing rework.

01Weeks 1 to 2

Design and map

Topology designed, compliance gaps fixed in Terraform before anything was provisioned.

02Week 3

Distribute state

Multi-region databases and EKS tuned for failover and rolling updates.

03Week 4

Operate

Observability, routine remediation and cost governance handed over.

The 4-week sprint. Compliance shaped the design from the first week.

Weeks 1 to 2: design and compliance mapping. The topology was designed for low-latency traffic between regions. At the same time, the proposed architecture was scanned while it was still code. Gaps in encryption at rest and IAM least privilege were found and fixed in the Terraform definitions before a single resource was provisioned. That's where the traditional route loses most of its time. A finding in a Terraform plan is a review comment:

resource "aws_rds_cluster" "workflows" {
  cluster_identifier = "workflows-primary"
  engine             = "aurora-postgresql"
  storage_encrypted  = true
  kms_key_id         = aws_kms_key.data.arn
  # ...
}

The same finding in a running account is a migration.

Week 3: distribute the state. Multi-region database clusters got automated failover, with replication lag tuned for consistency across regions. EKS got node scaling policies and rolling update strategies to meet the no-downtime requirement.

Week 4: operate and hand over. Observability that sorts incidents by severity, routine remediation for issues like hung pods and resource exhaustion, and cost governance with Savings Plan recommendations and resource scheduling to keep the new footprint on budget.

Automated remediation only works if the team knows exactly what runs on its own. CloudThinker acts within Auto Mode, so Diaflow decides what runs on its own (restarting a hung pod) and what waits for approval (a scaling change). Every action lands in the record, which a SOC 2 program uses twice: once for operations, and again as evidence.

One incident, end to end

In a distributed architecture, a symptom can start in the application layer (Kubernetes), the data layer (Postgres), the edge or the network. CloudThinker reads the monitoring Diaflow already has and correlates across all of them.

01Signals

Diaflow stack

  • GrafanaGrafana and PrometheusMetrics and alert rules
  • DatadogDatadogMonitors and logs
  • KubernetesKubernetes (EKS)Pods, nodes, rollouts
  • AWSAWSCloudFront, S3, databases, CloudTrail
  • GitHubGitHub and GitLabRecent changes and pipelines

Alerts arrive by webhook

02Investigation

CloudThinkerCloudThinker

  • CorrelateAll layers at once, one timeline
  • Test hypothesesEach reading checked against data
  • Root causeWith the evidence cited

Changes through Auto Mode

03Outcome

In the thread

  • Key findingsWhat fired and what it means
  • Recommended actionsFix plus prevention, with snippets
  • ResolutionApplied, verified, recorded

Fix and prevention together

Monitoring stays where it is. CloudThinker reads every layer and correlates them into one timeline.

This incident is real. The values below are illustrative.

01Report

Broken images

PNG assets served as text/xml instead of image/png.

02Readings

Three suspects

Wrong S3 metadata, a header rewrite at the edge, or a stale cache.

03Evidence

S3 vs edge

S3 says image/png. The edge serves an old cached copy.

04Fix

Invalidate

Cache invalidation created and confirmed complete.

05Prevent

Upload rule

Content-Type set at upload, with code snippets.

The CloudFront and S3 Content-Type incident, from report to prevention.

The report. Users saw broken image assets. PNG files failed to render because they arrived with Content-Type: text/xml instead of image/png.

The readings. A wrong header at the browser has three plausible sources, and each predicts something different:

  • The object in S3 has the wrong metadata. Then S3 itself returns text/xml.
  • Something at the edge rewrites the header. Then every object behind that cache behavior gets the wrong header, PNG or not.
  • CloudFront is serving a stale cached copy. Then the edge response is old (a high Age, a cache hit) and differs from what S3 returns now.

The checks. Compare what the edge serves with what the origin holds. The edge response:

curl -sI https://assets.example.com/img/workflow-card.png
# content-type: text/xml
# x-cache: Hit from cloudfront
# age: 172800

The object in S3:

aws s3api head-object --bucket example-assets --key img/workflow-card.png
# "ContentType": "image/png"

The finding. S3 had the correct metadata, but CloudFront was serving cached responses with the wrong header. Files had been uploaded without a Content-Type, fixed later in S3, and the cache stayed stale. CloudThinker named the distribution and tied the errors to those stale responses.

The fix. Create a cache invalidation and confirm it completed:

aws cloudfront create-invalidation \
  --distribution-id E2EXAMPLE123 \
  --paths "/img/*"

aws cloudfront get-invalidation \
  --distribution-id E2EXAMPLE123 \
  --id I2J0I21PCUYOIK
# "Status": "Completed"

The prevention. Set the Content-Type at upload time instead of fixing it afterwards:

aws s3 cp workflow-card.png s3://example-assets/img/workflow-card.png \
  --content-type image/png

An engineer working alone checks S3, then CloudFront, then the cache policies, then the upload pipeline, and only learns which layer mattered at the end. Here the comparison between S3 and the edge, which is the whole answer, came first.

Results

The figures below are from the Diaflow case study.

MeasureResult
Three-region infrastructure4 weeks
SOC 2, HIPAA and GDPR readiness3 weeks of those 4
Manual operations workDown 80%
Uptime99.9%

The longer-term change is who does the work. In a traditional model, the DevOps team grows with the platform. With the day-to-day infrastructure load handed to CloudThinker, Diaflow's conversations moved from "can the database handle this load?" to "how fast can we ingest this new dataset?"

"When we started, I honestly didn't believe this timeline was achievable. We were incredibly impressed by the flexibility and professional ability of the CloudThinker team to dive deep into our specific architecture. They didn't just resolve our technical debt; they helped us achieve the highest security standards and compliance in a very short time. It was a seamless experience from start to finish."

Jonathan Pham, Founder & CEO, Diaflow

Why this beats the traditional approach

Traditional routeDiaflow's route
Order of workBuild, then audit, then reopenArchitecture and compliance in parallel
Compliance gapsFound in live infrastructureFixed in Terraform before provisioning
Stateful servicesDiscovered late, where projects stallPlanned as its own week, with failover tuned
Day 2Lands on the developersRoutine remediation and cost governance
Incident debuggingOne layer at a time across consolesEdge and origin compared in one pass

The deeper difference is the order of the work. Compliance checked against code is a review comment, while compliance checked against live infrastructure is a migration. Running both in parallel against one codebase is what turned a months-long roadmap into a 4-week sprint.

Getting started

If you're facing the same mix of multi-region and compliance work, connect AWS with a read-only role and ask from chat:

"Users report images rendering broken from our CDN. Compare what CloudFront serves with what S3 holds for the affected objects and tell me which layer is wrong."

"Review our Terraform for the production account against SOC 2 controls. List encryption, IAM and logging gaps with the exact resource and the change that fixes it. Propose only, change nothing."

"Show me which resources in our second region are oversized compared with their last 30 days of use."

Next step: connect your AWS account and run the first prompt the next time an asset or header looks wrong. Resolve explains how an investigation runs from alert to verified fix.

Steve Tran, CTO, CloudThinker

About the author

ST

Steve Tran

Founder & CTO, CloudThinker

Steve Tran is the founder and CTO of CloudThinker, which builds Frontier Investigation Agents for production cloud operations. He writes about incident investigation, cloud cost, and the practices teams use to safely run autonomous AI agents against production AWS, GCP, and Azure environments.

agenticaiactivearchitecturesoc2complianteksoptimizationrootcauseanalysis