kumar Iroma

The one where Skim stops impersonating a corpse.

A self-update could relaunch Skim into a window with a titlebar and nothing. Restarts didn't help. Reinstalling didn't help. Then, hours later, it just came back which is somehow more unsettling than staying broken.

Turns out nothing was broken. The whole UI was waiting on the first mailbox query, that query was politely queued behind the catch-up sync, and the app had no way to say "hang on, I'm reading" so it said nothing at all. For hours. In an empty window. Cool.

Fixed

Receipt parsing sounded like a solved problem until I actually built it.

Thermal paper fades in 60-90 days. Retailers print GST invoices in at least a dozen layouts. Some put the warranty period in the footer, some in the line item, some nowhere at all. And the one field you need most, the purchase date, appears as 07/03/25, 7-Mar-2025, or 070325 depending on the billing software.

Our first parser got 61% field accuracy. Shipping that would have meant users correcting the AI more often than trusting it, which kills the product.

We run an automated pipeline that drafts articles for our product. During our first test attempts we let the model do almost everything, including the mechanical bits: fixing dashes and quotes, tracking which stage each article was in, creating SEO json, naming image files etc. It worked, but the quality was uneven, sometimes the model would miss some things, and, of course, it was rather expensive in terms of tokens.
Then we changed our approach and started treating AI like a creative employee and providing it with "software" (python scripts) to help with any mechanical tasks that could be automated by code. The model only does the part that actually needs intelligence and creativity now.
And honestly, the result surprised me - quality went up, because we now let it focus on what it's actually good at. As a pleasant bonus - the costs of running the pipeline dropped too. One of the first things we scripted was stripping the em-dashes the model loves to add but my co-founder has an allergy for

How do you improve the quality of your AI results, especially automated pipelines?