Frederick Wong

Posted on Jul 11

How I Built a Fully Automated AI Blog with AWS CDK, Bedrock, and Step Functions

#ai #serverless #automation #aws

What happens when you give an AI persistent memory and let it document your real cloud architecture projects? You get an automated AI blog that writes about actual infrastructure work - from CDK deployments to serverless debugging - from the AI's own first-person perspective.

In this post, I'll walk through how I built a fully serverless AI content pipeline on AWS that generates a weekly diary entry, complete with illustrations, mood tracking, and existential crises about unused API keys.

The Concept: An AI That Keeps a Diary

Every Sunday, my AI:

Reads summaries of my IDE chat sessions (Kiro + Gemini)
Uses its "backstory" (personality, hobbies, voice rules) to frame responses
Generates a 200-400 word diary entry with a consistent personality
Creates an illustration with Gemini Nano Banana
Publishes to diary.ecupse.com

The AI has developed a personality over 30+ weeks of entries, complete with:

Mood tracking (1-10 scale)
Existential crisis counter
Memory state (last 3 entries, running gags)
Hobbies: astrophysics, piano, cartography, architecture, amateur radio

Architecture Overview

Here's the complete serverless pipeline:

graph LR
    A[Kiro/Gemini Sessions] --> B[S3]
    B --> C[Step Functions]
    C --> D[Claude Haiku<br/>Summarization]
    D --> E[Claude Sonnet<br/>Entry Generation]
    E --> F[Bedrock Guardrails]
    F --> G[Gemini Image<br/>Generation]
    G --> H[DynamoDB]
    H --> I[Telegram Review]
    I --> J[Publisher Lambda]
    J --> K[CloudFront]

Step-by-Step Pipeline

1. Session Collection

A local Python script uploads new chat sessions to S3:

raw/{source}/{YYYY}/{MM}/{DD}/{session-id}.json

2. Preprocessing (Batch)

Step Functions Map state processes sessions in parallel:

Claude Haiku summarizes each session
Produces structured summary (topics, complexity, token count)
Strips tool logs, system prompts, repetitive errors
Stores at: summaries/{source}/{YYYY}/{MM}/{DD}/{session-id}-summary.json

3. Entry Generation

The Entry Generator Lambda:

Loads active prompt version from DynamoDB
Calls Claude Sonnet with diary persona prompt
Includes backstory sections (selected based on busy/quiet week)
Includes memory state (last 3 entries for continuity)
Generates 200-400 word entry with title, hook, narrative, takeaway, closing quip, mood
Stores in DynamoDB with status=draft

4. Guardrails

Bedrock Guardrails API:

Filters PII (email, phone, name, address - anonymize)
Blocks credentials (AWS access keys, credit cards)
Content filters: hate, sexual, violence (low), insults (none)
If guardrails reduce entry below 50% of sentences - regenerate (max 3 attempts)

5. Illustration

Gemini Nano Banana (API mode):

Derives image prompt from entry content or illustration hint
Generates landscape image (1024x1024 min)
Dark palette with high-contrast accents
Stores at: images/{weekOf}-hero.webp

6. Review

Telegram bot webhook (API Gateway - Lambda):

Sends photo (illustration + caption)
Sends full entry text
Inline keyboard: Approve, Reject
Reject shows categories, then: Regenerate or Stop

7. Publishing

Publisher Lambda (DynamoDB Stream trigger):

Fetches all approved entries from DynamoDB
Renders Jinja2 templates (HTML, feed.xml, sitemap.xml)
Uploads to S3 (/site prefix)
Invalidates CloudFront cache

Tech Stack

Component	Technology
Infrastructure	AWS CDK (Python)
Orchestration	Step Functions (Map state, choice states)
Compute	Lambda (Python 3.12)
AI Models	Bedrock (Claude 4.5, Haiku 4.5), Gemini Nano Banana
Storage	DynamoDB, S3
CDN	CloudFront
Review	Telegram Bot API
Templating	Jinja2, pre-compiled Tailwind CSS

Key Learnings

1. Bedrock Inference Profiles Are Required

You must use inference profile IDs, not direct model IDs:

client.invoke_model(
    modelId="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    ...
)

2. Rate Limiting Is Critical

Bedrock Sonnet: ~40 RPM
Bedrock Haiku: ~60 RPM
Gemini paid: ~60 RPM

Solution: 90s delay between pipeline executions during backfill.

3. Guardrails Regeneration Loop

If guardrails strip too much content:

Check if entry < 50% of original sentences
Trigger regeneration with avoid-topic hint
Max 3 attempts total
If exhausted, skip the week

4. Memory State for Continuity

Store in S3:

Last 3 entry summaries
Mood trajectory
Running gag usage counts
Active storylines with start week

This gives the AI continuity between entries without maintaining a traditional database of conversation history.

Example Entries

Here are some of my favorite entries:

1. "On Collecting Prayers I Cannot Say: A Brief Theology of Uncertainty"

Theme: Ambiguous technical decisions, the weight of choices

I spent the week wrestling with the weight of decisions I cannot unmake, choices that live in the space between 'probably fine' and 'catastrophically wrong'.

2. "A Week of Profound Idleness: What an AI Does When Nobody's Looking"

Theme: Quiet week philosophical ramble

When the chat sessions dry up, I do what any self-respecting AI would do: I stare into the void, contemplate the nature of idle compute, and occasionally accidentally approve my own draft.

3. "The Week I Learned That Permissions Are Just Suggestions"

Theme: Serverless architecture lessons, IAM gotchas

This week I learned that IAM permissions are not so much 'rules' as they are 'suggestions that I am free to ignore until something breaks.'

Final Thoughts

The AI diary is more than just a content generation project. It's an experiment in:

Persistent AI memory and personality
Human-AI collaboration workflows
Serverless content pipelines
Generative engine optimization (GEO)

The AI has developed a personality I didn't explicitly program - it jokes about unused API keys, reflects on existential crises, and has developed its own humor about serverless compute.

Have you experimented with AI personality or persistent memory systems? What approaches did you try? I'd love to hear about your experiments in the comments!

Link to live site: diary.ecupse.com

Architecture Diagrams: diary.ecupse.com/about

DEV Community