Where the sixteen seconds went
The first version took a median of over eighteen seconds per photo, which in practice means the seller puts the phone down and wanders off. Getting it to about two changed the behaviour entirely: people now photograph a whole pile of items in one go, because there is nothing to wait for.
Why the Finnish had to be real
Machine-translated product descriptions read as machine-translated and buyers notice. Getting natural Finnish out, rather than English labels stitched together or the item's own name repeated as its description, was most of the work.
An honest note on sovereignty
The model call runs on Vertex AI, physically in Google's Finland datacentre. The data stays on EU soil but the provider is American. That is a real exception to how we build everything else, and we would rather state it plainly than bury it.
- About 2 seconds per photo, down from 18 before we optimised it
- 19 out of 20 items land in the right category on their own
- Real, natural Finnish, not English labels stitched together
- Roughly $1 in model fees per 1,000 photos analysed
Also from workshop
Other things we have built.
KK Drive
A private, branded library where every store sees only the marketing it is actually allowed to use. Upload once and every eligible shop has it.
Read the full story → Daily driverToolbox
Half of software support is showing someone what to click. Screenshot, draw an arrow, save, attach, repeat five times. Toolbox collapses that into one keystroke.
Read the full story → In developmentSnapper
You spend the day dragging windows into place, and every time you unplug the monitor it all falls apart.
Read the full story →