We turn AI into your unfair advantage

Red Nose Day is one night, live on the BBC.

• It needed a payments platform that could not fail on live television.
• On live television there is no second take.

A new donation platform, built by four Armakuni engineers.

• It took 500 transactions a second.
• It was delivered on time and on budget.

Comic Relief broke £100 million for the first time.

• Zero downtime on the night.
• It was 2012, and it is where Armakuni began.

A bank's engineering teams had to move to cloud-native delivery.

• Four hundred engineers had to change how they work.

We changed how the teams build, not which tools they buy.

• Everything as code.
• Automated testing.
• A cloud-native delivery force that ships at speed.
• The practices are the AK Way, our own engineering practices.

40% faster time to market, across 400 engineers.

• The discipline stayed after the engagement ended.

Scale, matching and analytics all needed work at the same time.

• Thousands of hospitality employers, and millions of applications a year.
• Before each busy hiring window, engineers spent their time planning capacity.
• Search and data freshness had gaps.
• Recruiters handled much of the workflow by hand.

The platform rebuilt as a cloud-native foundation.

• Microservices on Amazon ECS and AWS Fargate, with serverless automation on AWS Lambda.
• Search on Amazon OpenSearch Service and Amazon ElastiCache.
• Resumes parsed automatically, candidates matched, and bulk job posting made asynchronous.
• Workflows joined to employers' applicant tracking systems.
• Analytics in one place, on AWS Glue and Amazon Athena.

Ten times the capacity: 600,000+ daily users at peak.

• Search answers 60% faster.
• Data indexing takes 45% less time.
• Recruiters do 80% fewer tasks by hand.
• A bulk job posting that used to take hours now finishes in minutes.

A delay was found after it had already cost someone.

• Dealers phoned operations to chase their vehicles.
• Shipping costs were steered by last week's report, because live data wasn't available yet.
• Adding a carrier meant a capacity-planning project.

Glovis America's first real-time tracking portal for dealers.

• Telemetry from IoT-enabled carriers, processed as a stream on AWS Lambda.
• A bridge to the Oracle data that already ran the business.
• Each dealer sees only its own vehicles: role-based access, and data kept apart by design.
• An alert goes out when a vehicle falls behind.

45% more vehicles delivered on time.

• Shipping costs down 15%.
• 40% more data capacity. A new carrier is now a change of configuration.
• 98% accuracy on data access restrictions.
• Far fewer calls from dealers asking for a status.

The platform was up 95% of the time.

• Hundreds of dealerships share it, so tenant isolation had to be hardened.
• The APIs had to be rebuilt for campaign peaks.
• Observability had to be rebuilt too.

Four pieces of work, on an event-driven foundation.

• A new security model that isolates every tenant.
• APIs rebuilt on an event-driven core: Amazon API Gateway, AWS Lambda, Amazon SQS and SNS.
• Lifecycle policies, auto scaling and right-sized compute.
• Infrastructure as code and automated deployment.
• Event-driven monitoring, with Amazon CloudWatch and AWS Security Hub.

99.9% uptime, held under peak campaign load.

• Zero cross-tenant exposure since the new model went live.
• Monthly AWS spend down 20%.
• New features reach market 30% sooner. A release takes days to line up, not weeks.
• 95% of API requests on time at peak.

Production kept failing, and nobody knew why.

• It had gone on for six to ten months.
• Each restart wiped the evidence.
• Nothing watched production, so every incident was a guess.
• About 1,000 policies and 2,000 quotes a day ran on two servers.

A DevOps Care pod: senior DevOps and full-stack engineers, for 60 hours.

• They kept the evidence from the last outage, and read it.
• The cause: Apache ran out of threads, with about 800 MySQL queries behind each request.
• The fix: connection pooling and session timeouts, chosen so a policy and its payment still commit together.
• Then eyes on it: Amazon RDS Performance Insights on the database, request tracing in the application.

An outage that ran for months, resolved in 60 hours.

• The platform is steadier, with no risk to a payment or a policy.
• The database and the application are no longer dark.
• The pod stays on a retainer for DevOps, database and full-stack work.

Every new customer was configured by hand.

• Bringing one on took six to seven months.
• The wait cost the company deals.
• To ask the data a question, staff needed SQL, a ticket and days.
• Billing problems surfaced in a payer's audit, after the fact.

Two pieces of work, side by side, in five to six weeks.

• Forms and checklists are built by generative AI on Amazon Bedrock. A person edits and approves each one.
• Staff ask the data in plain English and get an answer in seconds.
• Ten monitors watch the riskiest billing patterns as they happen.
• They catch overlapping visits, travel distances that cannot be real, and duplicate claims.

80 to 90% of the onboarding effort is automated.

• The rest is the person's check, kept on purpose.
• Six months becomes weeks.
• A compliance question gets its answer in seconds, with no ticket.
• The target: professional services cost down by 70% or more.

150 people listened to calls and labeled them by hand, every day.

• Tens of thousands of calls, for HVAC, plumbing and electrical contractors.
• A build of its own had not reached production quality.

A four-week proof of concept for AI call analysis.

• It classifies each call, labels how it ended and pulls out the outcome.
• Models benchmarked and fine-tuned on Amazon Nova.
• A total-cost analysis with a go or no-go framework, so the decision is a number.
• The data pipeline was audited and modernized first.

Production-ready in under two weeks of the four.

• The C-suite committed in the first half of the engagement.
• A fine-tuned Nova model opens a path to a lower cost per call at the same accuracy.
• Next came four proofs of concept at once. A prompt framework, missed revenue per call, forecasting and budget, alerts on anomalies.

A rushed move to Amazon Connect was failing a week before go-live.

• Thousands of customer contacts a week, across onshore and offshore centers.
• A big film release was the hardest test of the year.
• Coaching feedback took days to reach the people on the phones.

First we steadied the launch. Then we rebuilt it, in five pieces of work.

• AI agents on voice and chat: Amazon Connect with Amazon Nova Sonic, Amazon Lex and Bedrock AgentCore.
• They answer from live data on showtimes, refunds and loyalty.
• Contact Lens and Amazon CloudWatch put every call flow in view.
• Three chat workflows run with no person in the loop.

Calls are handled 30% faster across the whole contact center.

• A major release came and went with no emergency hiring and no all-hands weekend.
• Every call flow is visible, for the first time.
• Coaching feedback arrives the same day.
• A problem shows up in minutes, not days.

The contact center it had been left with was not performing.

• Frequent bugs, missing features, and releases nobody controlled.
• Outbound campaigns were fragile, and leads were recycled by hand.
• Supervisors could not see what was happening.
• Medicare disclaimers were not automated.

It chose us from several AWS partners. We began with a four-week assessment.

• It left a report on the current state, a risk plan, the architecture and a sized plan.
• Then a pod: CI/CD pipelines, pull-request reviews and release governance on Amazon Connect.
• A deployment model that keeps protected health information compliant, for the first time.
• A stateful layer runs outbound: each campaign follows a fixed lifecycle, with policy enforced.
• Contact Lens inside the call, live dashboards, structured Amazon CloudWatch monitoring.

600 people take calls on a stable, modern Amazon Connect.

• Releases are governed.
• Medicare disclaimers are automatic.
• Supervisors see a problem early enough to act on it.

The dealer built its own platform to run the business.

• It runs dealerships across Texas and Kansas City.
• The plan: sell it to other heavy-equipment dealers, as multi-tenant SaaS.
• We began in August 2025 with a 250-hour pod: stability, QA, CI/CD, monitoring and an AWS review.

In February 2026 the pod grew to 400 hours a month, to build the SaaS itself.

• Tenants are provisioned automatically, each with more than 150 database tables.
• Amazon Aurora Global Database, and disaster recovery.
• Federated single sign-on with Amazon Cognito, behind AWS WAF and Shield Advanced.
• APIs modernized, with Amazon ECS and AWS Lambda underneath.
• SOC 2 readiness, with a Type 1 audit next.
• AI workflow automation on Amazon Bedrock, and a portal for resellers.

60% more engineering capacity a month: 250 hours became 400.

• The first outside dealer to come on is in Kansas City.
• Signed under one CEO and expanded under the next, with no pause in delivery.
• Five roles flex month by month, with no fixed headcount.

The platform was manual, and locked to a vendor that was closing.

• A hard date: December 2025.
• Analytics were scattered.
• The model that ranks candidates needed steadying.
• The cloud foundation belonged to the vendor, not the firm.

A discovery funded by AWS, then five more contracts.

• Everything moved to AWS-native systems the firm owns and runs. The vendor's pipeline was replaced on Amazon EKS.
• Each new ranking model is checked against production, and rolls back by itself.
• An AI agent finds candidates and returns a ranked shortlist, with its reasons.
• Profiles update from recruiter calls and emails, and write back to the CRM.
• Revenue, margin and pipeline in Amazon QuickSight with Amazon Q.

Zero disruption through the shutdown.

• Every workload runs on systems the firm owns.
• Candidate ranking is monitored and quality-controlled, for the first time.
• The CEO and the board answer their own questions, with no analyst in between.
• All business data stays inside AWS.

The bill grew with the platform, and nobody could see why.

• Nine AWS accounts blurred who owned what, and who paid.
• No cost view went down to the container.
• Releases across 50+ dependent microservices were lined up by hand.
• Little live sight of data transfer, EKS workloads, backups and storage.

Cloud operations and FinOps, on a containerized platform.

• Amazon EKS standardized, with GitOps pipelines.
• Nine AWS accounts realigned into four, each with a clear owner.
• AWS CUDOS dashboards show cost by container.
• EKS workloads right-sized to what they really use.

25% lower cloud spend.

• Compute costs down 48%.
• 99.996% uptime, held through the migration and the cost work.
• Ownership, billing and governance are clear.

Nine platforms, and none of them joined up.

• The same person could be five separate records.
• Campaigns went to broad audiences, and spend was wasted on people unlikely to buy.
• A report took an analyst days, and every spending decision waited for it.

A customer data platform across all nine sources. We built it and we run it.

• A warehouse on Amazon Redshift, fed by pipelines on AWS Glue and AWS Lambda.
• Customers are now grouped by what they actually do, instead of by broad guesses.
• Segments flow to the email and ad platforms by themselves.
• Campaigns set off by events, across channels.
• Dashboards in Amazon QuickSight for marketing, eCommerce and events.

40% higher return on targeted campaigns.

• Participation in the Ruckclub loyalty program doubled.
• Customer engagement across channels is five times higher.
• Marketing, eCommerce and event teams get their answers 30% faster.
• One view of each customer, across all nine platforms.