AI Vision Inference

Using a real-world bookshelf scanner as the problem domain to explore serverless AI inference. AWS is the implementation platform throughout, but the patterns — event-driven architecture, structured output validation, agent frameworks — apply equally to GCP, Azure, and on-prem deployments.


Event-Driven Vision Inference with Bedrock, Pydantic, and Parquet

Part 1 of 2 in AI Vision Inference

How to integrate vision inference into serverless architectures without specialised ML infrastructure using AWS Bedrock, Pydantic validation, and Parquet storage.

— Written 17 March 2026 · Edited 27 March 2026 · Written by Benjamin Clark

Rebuilding the Bookshelf: Replacing the Manual Bedrock Chain with pydantic-ai

Part 2 of 2 in AI Vision Inference

How migrating from tightly-coupled Bedrock API calls to a pydantic-ai Agent decouples the AI integration from platform-specific code — making provider switching trivial, silent failures explicit, and tool integration possible.

— Written 21 March 2026 · Edited 27 March 2026 · Written by Benjamin Clark