Building an agentic app deployer with Amazon Bedrock and AWS Lambda

Every big company has the same graveyard of tiny tools that never got built: the one-off calculator, the intake form, the little dashboard someone's cousin cobbled together in Excel. PDI Technologies, a 40-year-old company serving over 200,000 fuel and convenience-retail locations worldwide, decided to stop letting those requests die in an engineering backlog. Its answer is PDI Brew, a system where an employee types out what they want in plain English and gets back a fully deployed, single-sign-on-protected web app running on AWS, no ticket, no Git repo, no waiting on a sprint.

The trick is splitting the job into two agents with very different jobs and very different levels of trust. A planning agent, built as a skill that runs inside an assistant like Claude, ChatGPT, or Claude Code, interviews the user and turns their request into a structured JSON manifest describing the app. For teams that need everything to stay inside AWS's walls, there's a second planning path that runs the same job through an Amazon Bedrock model invocation instead, logged entirely in AWS CloudTrail. Both roads produce the identical manifest, so nothing downstream cares which one was used.

That manifest lands on a provisioning agent, an AWS Lambda function that PDI deliberately kept out of the chat loop. Provisioning infrastructure is not the place for a model to improvise, so this half of the system is deterministic and fully auditable: it classifies the request as either a simple static app or a full-stack one needing its own database and Lambda function, then calls AWS and Microsoft Graph APIs like a checklist. Slow steps, like standing up a Microsoft 365 group, get handled by the Lambda calling a background copy of itself so the user isn't left staring at a spinner.

Most of these apps run on a shared, pre-warmed Lambda doing basic CRUD work, which keeps costs near zero when nothing's happening. Only apps that need something riskier, like sending email or hitting an outside API, graduate to their own dedicated Lambda and IAM role, and only after an admin signs off following automated code and permissions checks. AI features get the same cautious treatment: any app can tap into Amazon Bedrock, but every single call passes through a mandatory guardrail for PII and prompt injection, plus a twelve-counter budget system that can shut off spend at the level of a single user, app, or the whole company before it turns into a surprise bill. Admins also get a platform-wide kill switch that beats any app's own settings.

What's notable isn't that PDI let AI write app code — plenty of tools do that now. It's that they refused to let AI touch the actual provisioning decisions, keeping that half boring, logged, and reproducible on purpose.