This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend
What I Built
Derica is for my friend Amina, who sells rice, beans, garri and groundnut in a Nigerian market. She buys by the 50kg bag and sells by the derica, mudu and paint (three local tins). When her supplier raises the bag price, she has to redo the maths for every tin by hand. Her tin prices usually change late, and she earns less on every sale in between.
She pastes the supplier's WhatsApp message, for example "Rice 50kg now 78k". Derica reads it and shows what it understood. It then works out the new price for each tin, keeping her profit the same and rounding up to the next ₦50. One tap makes a price card picture she can share on WhatsApp.
Example: the bag goes from ₦78,000 to ₦82,000, and her ₦2,900 mudu becomes ₦3,050.
The AI only reads the message and never sets a price. Normal code does all the maths. If the AI reads something absurd, such as ₦15,000,000 for a bag, Derica refuses to show it.
Demo
Demo video:https://youtu.be/Md6PzFtvtso url:https://derica.onrender.com/
Code
https://github.com/JUICEWRLD998/Derica
How I Built It
- I trained a small AI model (Qwen3.5-4B) on Tinker with about 3,500 practice messages I generated. Training took about 19 minutes and cost under a dollar.
- Python with FastAPI serves the app. The page is plain HTML, CSS and JavaScript.
- The price maths uses exact fractions, so there are no rounding surprises.
- Render hosts it.
- Amina's real messages (18) were the base for the tests. I also wrote two harder sets of 39 and 29 messages and locked them before testing.
Why Does Open Innovation Matter?
Derica reads messy WhatsApp price messages, like "Beans no dey cheap again, 92k now". That job needs a model that knows Nigerian Pidgin and local measures such as the derica, mudu and paint. A general closed API would need a long prompt full of rules and examples on every call.
Open weights let me fine-tune a small model, Qwen3.5-4B, on about 3,500 practice messages. Open models also let me test honestly. I compared my model against two other models and a plain rules reader, with test sets locked before scoring.
My Agent Session
Prize Categories
- Tinker. I fine-tuned the model and served it through Tinker, and the app calls it live.
- Render. The app is deployed there.
Team Submissions: Solo builder- Mustapha Fadhlullah
Thanks for participating!
Top comments (0)