Building in public, week 13: the debt week nobody posts about

I'm building Convertify, a free image converter, in public over 52 weeks. It's a solo project on Rust + Axum + libvips with a Next.js frontend. This is week 13, and I want to write about it honestly because it was a week with no shiny new feature, and those weeks rarely get posts.

Here's what actually happened.

I chased a double-free instead of shipping

My "images to PDF" path had been crashing in production, intermittently, with a GObject assertion failure. Convert a few files, fine. Convert a few more, segfault. The worst kind of bug.

Root cause: a derived Clone on the libvips image type was doing a bitwise pointer copy of a C object without taking a reference. Two Rust values ended up owning one underlying GObject, and when both dropped, g_object_unref ran twice on the same thing. A textbook double-free, wearing a .clone() as a disguise.

// the bug

let copy = image.clone();   // copies the pointer, no g_object_ref



// the fix

let copy = ops::copy(&image)?;   // real new GObject with its own refcount

One line. A couple of evenings to be sure it was that line. I wrote the full story up as a separate dev.to article this week.

I fixed a UX dead-end I'd been ignoring

If you went over the 20 MB upload limit, the convert button disabled itself, and there was no obvious way out. You couldn't remove a single file from the list. People were quietly getting stuck.

So I rebuilt the file list: each file is a card with a remove button, there's a size-limit toast that fires when you cross the limit, and the dropzone goes red.

While refactoring the upload handler I also collapsed three separate file-handling paths (choose, drop, remove) into one gate. That killed a whole class of bugs where the three paths had drifted apart (drop wasn't updating the size map, re-selecting the same file silently did nothing, etc).

I audited my own content for honesty

This is the unglamorous one. I went through landing pages and removed claims that didn't match what the backend actually does, checking each against the real Rust code (metadata handling, compression details). I'd rather say less and be correct than pad pages with things that aren't true.

Funny side effect: an SEO audit I ran this week pointed at exactly this kind of honest, specific content as the thing that sets the site apart from templated competitors. The discipline turned out to be the moat.

The most useful thing I did wasn't code

I finally set up Bing Webmaster Tools. Took five minutes (you can import straight from Google Search Console).

It immediately showed me search intent Google had been hiding: around 25 different queries for "pdf to jpg at N dpi" (600, 300, 150), in multiple languages, where I'm already ranking on page one. People who need a specific print resolution. I had no idea that demand existed because Google's console never surfaced it. That's now the top candidate for next week's backend work.

Lesson: if you only watch Google, you're watching most of the market through one keyhole. Other search consoles are free and show you different things.

Numbers, honestly

Impressions and indexing are both at period highs (impressions peaked around 225/day this week). Clicks are still small and spiky, which is normal for a 3-month-old domain with almost no backlinks. Nothing regressed, despite a scary-looking week-over-week dip that turned out to be me comparing a 7-day peak against a 3-month total.

Threads is quietly my best referral source (real sessions every week), and ChatGPT keeps sending people too, which means I'm getting cited in LLM answers even though I haven't seen the AI crawlers hit me directly.

Next week is a real technical week

Backend debt: selectable TIFF compression, multipage TIFF, a CMYK path, and that DPI feature the Bing data justified. Plus refactoring my upload handler properly before I pile more on it, and turning my real benchmark data into a public study people can actually cite (which an audit suggested is a stronger link lever than guest posts).

No new converter pages. The lever this stretch is depth and correctness, not surface area.

If you're building something image-heavy, it's at convertifyapp.net. No account, files deleted 6 hours after conversion. Week 13 of 52 done.