I built an offline caption writer for small vendors, on my phone
My PC died the week this challenge started. So I built my project on the only computer I had left: my Redmi 13.
The problem
Small vendors in Nigeria post products on WhatsApp status every day. Writing a catchy caption each time is slow, and many can't spare data for online AI tools.
What I built
A caption writer that runs fully on a phone, with no internet needed after setup. I used PocketPal AI (open-source) with Qwen2.5-1.5B-Instruct (q4_k_m, about 1.1 GB), an open-weight model, plus a system prompt telling it to write short Nigerian English captions with price, delivery info, and a call to action.
Why open matters here
- It works offline, so no data is spent.
- It costs nothing per caption, so a vendor never pays an API bill.
- Product and price details never leave the phone.
- I can swap the model or change the prompt for any niche, like food vendors or thrift sellers.
What worked
For "Ankara tote bag, ₦8,500, delivery within Minna", the model wrote a clean caption with the correct price, emojis, and a "DM for more info" call to action. For "Homemade zobo, 50cl bottle, ₦700", it kept the exact price and bottle size. All of this ran on a budget phone at roughly 3 to 5 tokens per second.
What didn't
My first run was a mistake: the model treated my instructions as a request and invented prices I never gave it (500 naira, 300 naira). On two real tests it kept the product and exact price, but it dropped delivery details both times ("within Minna", "order a day before") and added a generic "Fast delivery" I never wrote. Adding "use all the details exactly" didn't fix it. A 1.5B model on a phone follows instructions loosely. Speed dropped from 5.5 to 3.2 tokens per second across the session, and replies took 10 to 26 seconds.
Honest note
This is a prompt-driven tool, not a trained model. I tested two products, not with a real vendor. Next I would try a larger 3B model, fine-tune on real vendor captions, and add Pidgin support
Top comments (0)