A veterinary practice, two clinics
Bookings, in the real diary
In the pilotAnswered 20 sample questions from the website
1,900messages a month, in production
A sick animal goes straight to a nurse
Take your AI pilot from demo to production.
Plenty of AI pilots work in the demo, then wait for months. Others are apps built with AI tools, live within weeks, that stall the day the first big customer sends its security questionnaire. Our forward deployed engineers work inside your business until your AI agents or your app run in production on your real data, safe for real customers, in about 90 days, with a decision gate for you at weeks 4, 8 and 12.
A one-day workshop with the engineers who’d do the work. AWS funding typically covers the workshop.
An AWS Premier Tier partner, one of about 300 among more than 130,000, validated by AWS for agentic AI consulting.
A sandcastle built for demo day, and a pier for every tide: AI agents at the gate, the lamps and the lookout, our engineers at the piles, and your owner taking the key.
Every pilot is built at low tide. Production brings the tide in.
On demo day the beach is empty: a dozen tidy examples, a clean export, and the person who built it at the keyboard.
Then the tide comes in. A real Monday brings every kind of case the demo skipped, all at once, and by Tuesday the pilot is switched off. It has waited ever since.
- The demo saw twelve cases. Monday brought 214, from photos of handwritten lists to “same as last week” and two orders in one email.
- A wrong answer cost more than no answer. Someone had to find every order it got wrong, so the team checked all of them, and the pilot saved nobody any time.
- The model was never the reason. What sank it sat around the model: words nobody had defined, like “same as last week”, a security review nobody had started, and no way to prove it would work at volume.
One order pilot, two daysA real Monday
orders entered right
What the demo never saw
Example
A pilot answers samples. Production answers everything that comes in.
In production, AI agents run the same idea on every message, order and contract, every day, and hand your people whatever needs a person. Here are six, from six kinds of business.
A small hotel group, four hotels
Answers from the real availability
In the pilotDrafted replies to 30 sample emails
86%of booking emails answered with no person
Every complaint goes to the duty manager
A law firm, 25 lawyers
Every contract against your checklist
In the pilotSummarized 10 sample contracts
140contracts a month, read against the checklist
A lawyer decides every marked clause
A courier firm
Answered from live tracking
In the pilotTracked 50 parcels in a test
3,400“where’s my parcel?” a week
A lost parcel goes to the depot’s supervisor
A heating and plumbing company
Jobs, into the real calendar
In the pilotSorted a spreadsheet of last year’s calls
310jobs booked a month
An emergency goes to the engineer on call
A car dealership
Every car on the lot
In the pilotAnswered 25 questions about one model
640questions a month, answered from the stock list
Finance questions go to the sales manager
Examples
Building it was never the hard part. Running it is.
The other way in is a pilot that’s already live. An app built with AI tools such as Lovable, Bolt, Replit, v0 or Cursor can have paying customers within weeks. Then the first big customer sends its security questionnaire. It asks where the data lives, who can read it, how the business would know if someone did, and how fast it would recover. Until the answers are good, the deal waits.
- AI tools put things where they’re easy to reach, so a database key often ends up in the browser’s code. Wiz found one AI-built app exposing 1.5 million API tokens through a key like that.
- Tables and storage often go live with no access rules, though each is one setting. A security researcher found data anyone could read or write in 170 of the 1,645 apps on one AI app builder’s showcase (CVE-2025-48757).
- The AI tool that built the app can still reach it. When building and live are one copy, one wrong command from it lands on the live database.
- Nobody wrote the code by hand, so nobody has the design in their head or can say what the next change will break. In Veracode’s 2025 tests, AI models wrote a security flaw into 45% of their coding tasks.
Their security questionnaireToday
questions
answered well
Example
Each app runs part of a business. Each had something left open.
Most were built in weeks by someone who isn’t a developer. Our AI agents find what’s open, our engineers close the urgent gaps first, and the people who built the app keep building with AI. Six examples, from six kinds of business.
Bookings and treatment notes, at a physiotherapy practice with three clinics
Anyone with the key in the browser could read every booking.
2,400patients’ bookings behind access rules in week one
The practice manager still changes the app with AI, through a review.
Quotes from a photo of the job, at a building firm
Its AI model key sat in the browser, so anyone could run up its bill.
340quotes a month, with the key replaced and kept on the server
The estimator who built it still runs it.
A portal for clients’ tax papers, at an accounting firm
Clients’ uploads sat in storage anyone with a link could open.
1,100clients’ files in closed storage, with links that expire
Clients still upload from their phones, as before.
The members’ app, at a gym group with five sites
The AI tool that built it could still change the live database.
6,000members, with the AI tool kept out of production
Changes reach members only after staging.
Trade orders, at a food wholesaler
Backups ran every night, and nobody had ever restored one.
900trade customers, with a restore tested and timed
Orders keep coming in overnight, as before.
Matching candidates to roles, at a recruitment agency
Its biggest client sent 140 questions, and most of the answers were already in its code and settings.
140questions answered, most of them by the design itself
The founder answers the next questionnaire in an afternoon.
Examples
Three gates between the demo and production. Your owner signs each one.
Nobody photographs a pier’s piles, but they’re why it still stands. Ours are three phases, each ending in a gate. A Forward Deployed Lead runs them, an AWS solutions architect who builds, with engineers for your data, the AI agents and the tests. We fix the metric and the pass mark before we build anything. On an AI-built app, AI agents do the reading, the scanning and the testing, our engineers make every call, and old and new run side by side until the new one is proven.
- 1
Weeks 1 to 4: find out where it really stands.
We run your pilot on a sample of your real cases first, so its score is a number, not a memory of the demo. We write down what your business means by its words, so the AI agent never guesses what “same as last week” means. Data access and the security plan start in week 1, while there’s time to act on what they find.
On an app built with AI tools, week 1 closes what’s open first.
- Our AI agents scan the code and the live app for exposed keys, tables and storage with no access rules, and missing access checks.
- They look for packages that don’t exist, too. AI tools invent the names, and attackers register them.
- We treat every key that ever reached the browser as public. We move it to the server and replace it, so the old one stops working.
- We put access rules on every table and storage bucket, then test them by signing in as a stranger and as the wrong customer.
- Your AI coding tools lose their way into live data. We restore a backup, time it and write the time down.
Then we read the app before changing it.
- Our AI agents read it into specs: what each screen and each rule does today.
- Characterization tests pin that behavior, right or wrong, before each piece moves. It’s how AK Evolve rebuilds old systems.
- We move the access rules out of the browser and the database into an API, where each rule is written down and has a test of its own.
- The database keeps its own rules as a second lock.
At gate 1, in week 4, your owner signs the scorecard (the metric, today’s baseline and the pass mark) and the path through data and security. Go, or stop and keep what we found.
Why:the pilot is judged on a number rather than on how the demo felt, and the security questions come up in week 1, not in month nine, when they stop everything. On an AI-built app, a key from the browser is all a stranger needs to read your customers’ data, and one wrong command from the AI tool that built the app can wipe the live database in the middle of a fix. If nobody knows what the app does today, a move can quietly change what it does for a customer.
- 2
Weeks 5 to 8: build it into your systems, in your AWS account.
Plain code handles the rules, the AI does the reading, and a person does the rest. Code checks every order against the order system first: the product exists, the quantity makes sense, the account is open. Whatever fails a check, or whatever the AI isn’t sure of, goes to a named person with a note saying why, and so does anything with weight, like a new customer’s credit.
Every call goes through one governed gateway on Amazon Bedrock AgentCore, with one narrow, approved connection (an MCP server) for each system. Each call is metered and tied to the person the AI agent acts for, and the AI agent has its own identity and only that person’s rights, checked outside the AI.
Each action carries its own key, so a timeout never puts an order in twice. A change to a prompt, a model or a tool reruns the whole test set: your real cases and past failures, grown past 100, with a sealed set nobody tunes against and critical cases that must pass every run.
An app built with AI tools moves into your AWS account in these weeks, services first and the database last.
- Sign-in goes to Amazon Cognito, files to Amazon S3 and functions to AWS Lambda or AWS Fargate. The app is served from Amazon CloudFront or AWS Amplify Hosting, and secrets live in AWS Secrets Manager.
- Customers move to Cognito the next time they sign in, through its user migration step, so nobody has to reset a password.
- Model calls go behind an API on Amazon Bedrock, with guardrails, instead of a key in the browser.
- Postgres moves to Amazon Aurora PostgreSQL last, once nothing else depends on the old database.
- Until the cutover, AWS Database Migration Service copies every change across, so old and new hold the same rows. Each piece is proven before the next.
- The cutover keeps a window to roll back. If something is wrong, customers go back to the old app while we fix it.
One physiotherapy practice’s booking app moved this way over eight weeks, six pieces one at a time, while its 2,400 patients kept booking (an example).
At gate 2, in week 8, every step works, from what comes in to what goes into your systems, in your AWS account and on your real data.
Why:a confident wrong answer never goes straight into your systems, and the same order is never entered twice. On an AI-built app, there’s no big-bang weekend for your customers to notice, and no data left behind on a platform you don’t control.
- 3
Weeks 9 to 13: prove it at volume, then ship it.
It runs beside your team on every real case, scored against what they did, with cost per task measured rather than estimated. Before go-live, we name the failure modes with you and set the human checkpoints.
One branch or team goes first. Its answers are checked on live traffic, with a line agreed in advance that rolls it back and a kill switch we have rehearsed. Anything it already entered is checked against your systems and put right, because a rollback can’t take back an order that has gone in. We trace each step, and each failure becomes a new test.
Whoever runs it day to day passes a drill without the people who built it: deploying a change, restoring from a backup and being on call. That’s your people, with runbooks, alerts and a named owner, and our engineers stay on call beside them until they pass. Or it’s our managed service, AK Run, which looks after incidents, cost and drift.
An app built with AI tools keeps its AI, and we put it on rails.
- Building, testing and production get separate AWS accounts. Your AI tools and our AI agents work in the first two, never in production.
- A pipeline builds, tests, scans and deploys as code, with review gates.
- Coding standards and steering files sit in the repo, so whoever built the app keeps building with AI at the same speed.
- We design in encryption, audit logs, least-privilege access and tested recovery, so the evidence a SOC 2 or HIPAA auditor asks for is already there.
At gate 3, in week 12, go-live is decided on the evidence, once your people have accepted the results.
Why:nobody finds the same problem on the same morning, and the process keeps running when the person who built it moves on. On an AI-built app, the next AI-written change can’t reach customers unchecked, and the next questionnaire doesn’t stall the next deal.
The same order pilot, week by weekIn production at week 13
right first time,
at week 13
205 of 214 on a real Monday, so Rosa and her team read 9 notes, not 214 orders, and nothing goes in twice.
Example
What a client said
…a very successful delivery of MVP of our Call Intelligence. … All clients from [one of our brands] have now been moved from manual tagging to Call Intelligence. This represents about 500 clients. … We have also started to see the ROI we were expecting.
Dinesh D., CTO, Blue Corona
Armakuni were the engineering partner who made the AWS migration work. They came in with a clear plan, ran it cleanly, and the team that came out the other side had levelled up.
Marcus East, Former CTO, Experian
One day to find out if your pilot can make it.
The workshop is How to take your AI pilot to production, a day with the engineers who’d do the work. Bring the pilot and 20 cases with known answers. AWS funding typically covers the workshop.
- In the morning, we score your pilot on those cases and map what it reaches, with whose permissions, what a task would cost at real volume and what happens when it’s wrong.
- In the afternoon, we design the way to production: the metric and pass mark, the path through data and security, and the three gates.
- You leave with a findings report, a target architecture and a costed roadmap: what to keep, fix and add, and our verdict. It’s yours to act on, with us or without us.
- Then the work is an outcome rather than hours: one process in production, measured against the pass mark you signed.
If you bring an AI-built app
The day is How to secure and scale your AI-built app on AWS. Bring whoever built the app and whoever will own it. The urgent fixes start in the room.
- Before the day, our AI agents scan the app with the read-only access you give us. They map every platform it leans on: sign-in, storage, functions, hosting and model APIs.
- In the morning we walk through the findings with you: what data the app holds and where, what breaks first at ten times your customers, and what your first big customer will ask that it can’t answer today. We start the most urgent fixes in the room. On one booking app we fixed two of its seven findings before the day ended and started three more (an example).
- In the afternoon we design where the app lands in your AWS account, sized to it, and cost the move.
- You leave with the findings ranked by risk, a target architecture and a costed, phased plan, yours to act on with us or without us.
What we’ll tell you not to do
- Don’t plan on beating your security review. Nobody ships faster than your security team approves, so if data access or the review won’t clear inside 90 days, we say so in week 1 and plan around it.
- Don’t use AI agents where rules will do. If plain code or one model call does the job, that’s what we’ll recommend.
- Don’t start without an owner. If nobody owns the process, or the data can’t be reached in about four weeks, you’ll hear it before you spend anything.
- Don’t move an app that’s fine where it is. Some only need three settings changed, and some of what we find costs nothing to fix.
- Don’t keep an app nobody uses. We’ll say which to switch off, even when the answer isn’t a project for us.
- Don’t take the AI tools away from the people who built the app. The speed is why it exists, so their changes go through a review and a pipeline instead.
What stays in your AWS accountWhen our engineers step back
in your AWS account,
made week by week
We’ve taken AI past the demo before.
70%
of the proofs of concept we scope reach production
500+
customer launches on AWS
1,300+
engineers, across the UK, US, UAE, India and Canada
16
AWS service delivery designations, and the Migration and Modernization Competency
Ryno Strategic Solutions
Marketing for home services companies: heating and cooling, plumbing, electrical
More than 150 people analyzed tens of thousands of inbound calls by hand every day, and the company’s own attempt had not reached production quality.
We ran a four-week proof of concept for AI call analysis: classifying calls, labeling how each ended and pulling out the outcome. Amazon Nova was benchmarked and fine-tuned, with a total-cost analysis and a go or no-go framework.
- Production-ready in under two weeks of the four
- The C-suite committed in the first half of the engagement
- The target is to take that manual work off their hands. AWS funded $25,000 of the work.
DigitalStrike
A multi-tenant marketing platform for hundreds of car dealerships
- Before
- It needed enterprise-grade reliability and hardened isolation between its dealerships, and its API performance and observability rebuilt.
- What we built
- A new multi-tenant security model, deployment automation and infrastructure as code, faster APIs, cost work and event-driven observability.
- Zero cross-tenant exposure since the new security model went live
- Uptime 99.9%, up from 95%, under peak campaign load
- Carmakers’ compliance reviews are answered from the architecture
- Monthly AWS spend down 20%
A government case management platform
Case management software for government agencies and non-profits across counties and states
Onboarding a new client took six to seven months, configured by hand, and deals were lost on the timeline.
We ran two fixed-scope pieces of work in parallel, in five to six weeks. One automates form and checklist onboarding with generative AI, followed by a person’s edit-and-approval step. The other gives plain-English analytics, with ten monitors for the riskiest billing cases.
- 80 to 90% of form onboarding effort automated; the rest is the person’s step
- Onboarding from about six months to weeks
- Compliance questions answered in seconds. Largely funded by AWS.
An executive search firm
A global C-suite and board-level search firm
- Before
- A manual platform locked to one vendor, and the vendor shut down in December 2025.
- What we built
- A discovery fully funded by AWS opened six pieces of work. We modernized the infrastructure, replaced its pipelines, built an AI agent that returns a ranked, reasoned shortlist, and added analytics. Each new ranking model is checked against production, with an automatic rollback.
- Zero disruption through the shutdown
- Every workload on AWS-native systems the firm owns and runs
- All of its business data stays inside AWS
What owners say before they start.
Got another? Bring it to the workshop.
Our own engineers can finish it.
They may, and the workshop will tell you. What usually holds a team up is the part a demo never needed: a test set with known answers that every change has to pass, the gateway and the checks between the AI and your systems, and the rollback, the traces and the on-call once it’s live.
Our engineers bring that part with them, and whoever will own it on your side works beside them from week 1.
Can you rescue the pilot we have, even on a tool we bought?
Often, yes, and it’s a good place to start. We test what you have on your real cases first. The parts that hold up are kept, and so is what your team learned building it.
What tends to need building is the part a demo never needed: how it reaches your systems, how it’s tested and how it’s watched. Everything we build stays in your own AWS account, with full source. If it isn’t worth finishing, we say so at week 4.
Our platform handles security.
The platform secures itself, and it does that well. The rules on your own tables and storage are still yours to set, and that’s where the public leaks came from.
The scan before the workshop shows which of yours are open. Often it’s a few settings, and some of what we find costs nothing to fix.
We’ll rebuild it properly later.
That may be the right plan, and the workshop will say so if it is. What’s open today can’t wait for it, so we do that part now. We treat keys in the browser as public and replace them, put access rules on every table and storage bucket, and restore a backup so you know it works.
A rebuild later starts from specs and tests of what the app does today, which helps either way.
We don’t have the budget for another project.
AWS funding typically covers the workshop. AWS has funded part of the work for several of our clients, and the workshop tells you whether yours could qualify.
The work is one process, scoped at the workshop, with a go-or-stop at week 4. If the answer is stop, you keep the map, the baseline and the scorecard.
On an AI-built app, many of the urgent fixes cost nothing but an afternoon. The plan is phased and costed, so you can stop after any phase and keep what’s done.
We don’t want to lose our speed.
You won’t. Nobody takes the AI tools away; they just stop shipping to production unchecked.
Building and testing happen in their own AWS accounts, where the AI tools can go as fast as they like. Production takes changes only from the pipeline, after the tests and the review gate.
Bring the pilot that never went live, or the app your customers already use.
Spend one day with the engineers who’d do the work, on your pilot and 20 of your real cases, or on your app. You leave knowing where it stands, the metric and pass mark that would prove it, and whether it’s worth taking further. Bring an app built with AI tools and you also leave knowing what’s open, what’s already fixed, what to tell your biggest customer, where the app should run and what the move would cost. AWS funding typically covers the workshop.