craigthompson.dev

$ cat case-studies/ai/platform.md

AI where it earns its keep

// ai platform · mastra + bedrock · 2024 → 2026

AI where it earns its keep

I helped introduce AI into the product: an agent service built with , running models through in its own isolated AWS account, with an that runs in CI.

How AI reaches the product
product
features that call AI
in the core platform
  1. every change:
  2. PR to the AI service
  3. eval suite in CI
  4. ships only if it passes
# in practice
  • AI features are built as agents and workflows in TypeScript, so any engineer on the team can work on them.
  • The AI service has its own account, with its own permissions and its own bill, separate from the core platform.
  • Changes to the AI service are checked by its eval suite in CI before they reach customers.

# terms used

Mastra
An open-source TypeScript framework for building AI agents and workflows.
Amazon Bedrock
AWS's service for using foundation AI models through an API, inside your own AWS account.
Eval suite
Automated tests for AI behaviour: set inputs checked against expected qualities, run on every change to catch regressions.

$ ls case-studies/

Search and jump