AI-Powered Daily Blog

Problem

Writing consistently is hard to sustain alongside other work. I wanted a blog about AI engineering and developer tools that publishes on a schedule, and I wanted to treat it as a real production system: automated, observable, and safe to leave running.

My role

Sole engineer. I designed the agent pipeline, the publishing flow, and the site at blog.haheng.au.

Architecture

1. 01

   Scheduler

   Triggers a run on a fixed schedule

3. 02

   Research agent

   Finds and summarises a timely topic

5. 03

   Writing agent

   Drafts the post in the blog's voice

7. 04

   Review step

   Checks quality and guardrails

9. 05

   Publish

   Post, cover image, RSS feed

The blog itself is a Next.js site with an RSS feed per language, which is also what this portfolio reads to show the latest posts.

Agent design

• Separate roles. Research, writing, and review are separate steps, so each prompt stays focused and each output can be checked on its own.
• Structured hand-offs. Each step passes structured output to the next instead of free text, which makes failures easier to detect.
• Bilingual output. Posts are published in English and Indonesian.

Key decisions

• Treat it like a service, not a script. Scheduled runs, logs, and clear failure states, so a bad run doesn't silently publish.
• Guardrails before publish. A review step sits between generation and publishing.
• Static output. Posts are published as static pages, so the site stays fast and cheap to host.

Evaluation & metrics

The pipeline publishes new posts on blog.haheng.au on a regular schedule.

Limitations & safety

• LLMs can state things confidently that are wrong. Posts are framed around sources and practical takeaways, and the review step is there to catch obvious errors.
• Automated content needs a kill switch. Runs can be paused without touching the published site.

Tech stack

AI agents · LLM APIs · Next.js · Scheduled jobs · RSS

Lessons

• Most of the engineering in an agent system is around the model: scheduling, validation, retries, and deciding when not to publish.
• Small, single-purpose agent steps are much easier to debug than one large prompt.