~/blog

An AI meme generator on AWS Bedrock

#aws #bedrock #serverless #ai

$ curl $WEBSITE_URL/generate -d '{"topic":"monday standups"}'
{ "caption": "..." }   # straight from Claude, via Bedrock

The Pooh Meme Generator is a silly app with a serious backend: you give it a topic, a model writes the caption, and you get a Winnie-the-Pooh meme to download or share. The reason it exists is the question behind it — how do you put generative AI into a real, deployable app without running any AI infrastructure yourself? Live demo →

The architecture

[browser] ──prompt──▶ [Lambda] ──invoke──▶ [Bedrock: Claude 3 Haiku]
                          │                          │
                          └────────── caption ◀──────┘

A static frontend on S3, a Python Lambda as the glue, and Amazon Bedrock doing the language generation. The whole thing is serverless and provisioned with Terraform.

Why Bedrock

There’s no model to host, no GPU instance to keep warm, no inference server to babysit. Bedrock exposes the model behind an API and bills per token. From the Lambda it’s a single InvokeModel call, and access is governed by IAM like any other AWS service — the function gets a tightly scoped policy and nothing more.

The model is pinned in Terraform as a variable:

TF_VAR_model_id = "anthropic.claude-3-haiku-20240307-v1:0"

I chose Claude 3 Haiku on purpose: captions are short and need to come back fast, so the smallest, cheapest, lowest-latency model is the right tool. Swapping models later is a one-line change — the infra doesn’t care which model ID it points at.

Tuning the Lambda

TF_VAR_lambda_memory_size = 512
TF_VAR_lambda_timeout     = 10

512 MB is plenty for a function whose real work is a network call to Bedrock, and a 10-second timeout comfortably covers generation while still failing fast if something hangs. Right-sizing here is the difference between a function that costs cents and one that quietly wastes money.

What it demonstrates

  • Generative AI on managed infrastructure — Bedrock instead of self-hosted models.
  • Serverless integration — Lambda calling an AI service, secured with IAM.
  • Infrastructure as Code — the model, the function, and its limits are all declared in Terraform and shipped through CI.

It’s a meme generator. It’s also a working pattern for adding AI features to a cloud app without standing up a single GPU.

Code: github.com/Breno30/pooh-meme-generator

$ cd ..