Inspiration
A property manager asked me for this. One of their buildings has more than 200 apartments, and in December the concierge desk is full of parcels. A resident asks for a parcel and someone searches the shelves by hand. When a resident says it never arrived, nobody knows what happened. And they don't want residents' names on someone else's cloud.
What it does
- Check-in: photo of the label, the Skill finds the resident and proposes a shelf. One approval for the whole batch.
- Find and pickup: "what is there for 12B?" gives parcels and shelves; the handover is logged with time and concierge on shift.
- Courier check: photo of the delivery sheet, compared with the register: "52 on the sheet, 51 registered, the DHL for 7A is missing".
- Daily review: parcels waiting too long, labels with no match.
How I built it
A SKILL.md following the Uliya standard, delegating to photo-agent, document-agent and file-agent. Before writing it, I tested the risky parts on my mini PC with a small local vision model and invented labels.
Challenges
The small model read every label fine, but when matching it sometimes made up an "alternative name" and gave a parcel to the wrong person. So now every match has to quote the register row word for word, or the parcel goes to the concierge. After that, zero parcels to the wrong resident in all runs. A misspelled name sometimes ends up with the concierge, that's ok, a question is cheaper than a lost parcel.
What's next
Testing on a real NAS, with creased labels and handwritten courier sheets. All test data is invented.
Log in or sign up for Devpost to join the conversation.