<img height="1" width="1" style="display:none;" alt="" src="https://px.ads.linkedin.com/collect/?pid=8744922&amp;fmt=gif">
Skip to content
AI Adoption Acceleration

Forward deployed engineering, with the patterns to prove it.

Cloudsoft engineers work inside your team to get AI out of pilots and into production, as we already do at some of the world’s largest technology companies.

In partnership with AWS, we’ve turned that work into repeatable patterns that make AI reliable, even for organisations in regulated industries.

Talk to an engineer →

Where AI gets stuck

Illustration: 33 dots, 4 highlighted. For every 33 AI pilots, only 4 reach production.

Most AI projects stall at proof of concept.

For every 33 that companies launch a pilot, only 4 reach production.

Source: IDC, via CIO.com

On its own, a model makes small but crucial errors, and for processes in regulated industries a small error is still a failed check. “The system said so” isn’t an answer an auditor accepts.

What we do is ensure the AI is reliable and auditable.

What we’ve built

AI in production for a global healthcare technology company, a mortgage technology provider and a major cloud provider.

Healthcare: thousands of pages of regulation, turned into action people can rely on

Its customers have to check safety data against US federal regulations and the rules of every state they operate in. The rules run to thousands of pages, they’re written for lawyers, and they change.

We built a pipeline on Amazon Bedrock that reads federal and state regulations, turns each section into plain-language guidance, and maps the state rules against the federal ones. The company’s compliance experts review every piece of guidance, and their edits are applied exactly as they make them.

Compliance checks pair Claude on Amazon Bedrock with AWS Automated Reasoning checks, so the parts that must be exact are deterministic and reproducible, and the parts that require LLM judgement are tightly bound and labelled to maximise reliability, with both easily inspectable when proof is needed.

Mortgage: from proof of concept to production

Before Cloudsoft a proof of concept showed AI could check loans against hundreds of lending guidelines. What it couldn’t do was show why a loan passed or failed. For a lender, compliance has to be proven, not just asserted.

The steps to prove compliance

  1. At design time
    • AI and human experts collaborate to formalise and review the requirements
    • AWS Automated Reasoning checks prove that rules don’t contradict and aren’t redundant
  2. At evaluation time
    • AWS Automated Reasoning and deterministic code run the checks where logic is exact.
    • Claude on Amazon Bedrock handles individual checks that need judgement, such as analysing explanations or inspecting photos, and those results are flagged as lower confidence.
  3. The evaluation report
    • Gives an overall verdict and confidence, highlighting the items that need attention
    • Shows every result traced back to the rule and the source wording behind it, coupled with provenance of specific data used for each check

Cloud infrastructure: AI engineering at cloud-provider scale

Many enterprises won’t rely on a cloud service until their infrastructure and compliance tools can see and manage it. At one of the world’s largest technology companies, thousands of resource types were missing that coverage. Building each one by hand took weeks of specialist engineering, and a mistake could cause an outage for customers.

We built an AI engineering system on Amazon Bedrock that does the build work in the background. Nothing it produces is promoted until it passes up to 38 automated checks that the AI doesn’t control: it can’t mark its own homework. Engineers step in only when a check fails.

The pattern behind it

Every engagement is different, but the principle isn’t: use AI, but don’t trust it alone. AI does the work it’s good at, and something other than the AI checks the result.

In technology, that’s automated checks on every build. In healthcare, it’s experts approving every piece of guidance. In mortgage compliance, it’s formal proof.

A closer look: AI you can trust in regulatory environments

We’ve built a repeatable reference architecture on AWS for checking work in regulated processes. It runs serverless on AWS, on Bedrock, AgentCore, Lambda, DynamoDB and S3, so is capable of handling very large scale in parallel for speed. Runs are retained and can be replayed on demand, and reports can be digitally signed and downloaded as an attestation.

Architecture diagram of the V8 Rubric Engine on AWS: authors build and publish rubrics, and every submitted artifact is evaluated against them using code, Automated Reasoning and Claude.

1. Author

AI drafts the rules from policies and documents, and experts review and approve them. Automated Reasoning finds rules that contradict each other, can never trigger, or repeat another. Each published version is frozen, so there’s always an audit trail.

2. Check

Every case runs through code for exact logic, Automated Reasoning for formal proof, and an AI model, from any provider, only where judgement is needed.

3. Report

Each result comes with its evidence: which rule, which data, and why. Reports can go out as a signed PDF that an auditor will accept.

How we work

“Forward deployed” is fast becoming a label every AI partner uses. For us, it’s simply how we delivered the customer stories above.

Our engineers work inside your team, in your repos, your stand-ups and your release process. We’ve done it inside one of the world’s largest cloud providers, inside regulated businesses, and inside product teams shipping to thousands of customers.

What our engineers bring is harder to copy:

✓Patterns already proven in production.
✓Which parts of a process AI should handle and which it shouldn’t.
✓How to check every result with something other than the AI.
✓Where experts need to approve, and how to do that without slowing everything down.
✓How to hand over a system your team can run and change.

You’re not engaging Cloudsoft to learn on your project. We’ve already done that work, and we bring it with us.

We build on Amazon Bedrock, often with Anthropic’s Claude models, and work in partnership with AWS.

Technologies we build with

  • Amazon Bedrock
  • Anthropic Claude
  • Kiro
  • Amazon Bedrock AgentCore
  • AWS Automated Reasoning checks
  • Model Context Protocol (MCP)
  • AWS Step Functions
  • Amazon DynamoDB
  • AWS Lambda
  • Amazon S3

Need to trust and prove your results?

Spend 30 minutes with one of our engineers. Tell us where you’re stuck, and we’ll tell you what it would take to get it into production, and whether AWS funding could help.