Writing about tech, personal growth, and a little love story

This article explains how we designed an AI-powered question generator system for UTBK preparation with a strict focus on keeping operational cost near zero. The system relies on Google Gemini 2.5 Flash as the primary model, but the real cost efficiency comes from workflow design: controlled batching, throttling, deduplication, lifecycle-based content management, and drip publishing. Rather than optimizing for maximum AI usage, we optimize for sustainable daily generation within free-tier constraints while keeping the system aligned with the mission of providing accessible education through terasbelajarasik.web.id/bank-soal.

Running a learning platform like terasbelajarasik.web.id comes with an interesting challenge: it’s a free learning platform, which means the hosting costs should ideally be free too. And somehow, I’ve managed to pull that off — my monthly cloud bill is exactly $0, despite running three separate apps across multiple platforms.

In my previous post, I shared how I run my own server using an Android TV box with Armbian and Cloudflare Zero Trust. While this setup works great for serving applications to the internet, I quickly ran into another challenge: how do I deploy updates from GitHub Actions?

Deploying a server used to be a straightforward decision for me: pick a hosting provider like Vercel, Fly.io, or a VPS service, and get going. But as I dove deeper into my projects, I realized there were limitations and costs that made me rethink this approach. In this post, I want to share how I ended up running my own server without relying on cloud providers or a static public IP address.