This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
Lazy Meal Diary: an offline cooking helper for students, on my phone
My PC is still broken. Last week's post was about the day it died, and it's still not fixed, so this build also had to run on the only computer I have left: my Redmi 13.
The problem
Student cooking has the same loop every week. You cook something decent on Tuesday, forget it by Friday, and stand in front of the same few ingredients with no idea what to make. I wanted a diary I could talk to in one line, one that remembers what I cooked and suggests something similar later.
What I built
- App: PocketPal AI (open-source), running fully on the phone
- Model: Qwen2.5-1.5B-Instruct (q4_k_m, about 1.1 GB), an open-weight model
- A system prompt that makes it a meal diary for a Nigerian student with one burner and a small budget
- One dedicated chat for the whole week
How the "memory" works
The model has no memory of its own. PocketPal keeps the chat, and the model re-reads the whole chat every time I send a message. So as long as I keep one chat open, Tuesday's meal is still in front of it on Friday.
Why open matters here
- It works offline, so no data is spent.
- My eating habits and budget stay on my phone and not on someone else's server.
- It costs nothing per message, so I never pay for it.
- I can change the model, the prompt, and the generation settings.
What worked
I logged "Tuesday: egg and indomie with tomato" and "Wednesday: beans and garri". On Friday I asked for something similar to Tuesday's meal, and the model correctly named "egg and indomie with tomato". Replies came in about 2 seconds after the model loaded, at around 6 tokens per second.
What didn't
- A repetition loop. My first attempt made the model write a random recipe and then repeat "1/4 cup chopped cilantro" over and over. I fixed it by lowering temperature, raising the repeat penalty, and capping the reply length.
- It ignored my "one short line" rule. When I logged a meal, it often answered with a recipe instead of a short confirmation.
- It ignored my new ingredients. On Friday I said I had eggs, rice, and pepper. It recalled Tuesday's meal but used none of those ingredients, and it invented "egg noodles".
A 1.5B model on a phone can recall what's in the chat, but it doesn't follow detailed instructions well.
Honest note
This is a prompt-driven tool, not a trained model, and a long chat could overload a small model's context. I tested four messages in one chat and not with another student. Next I'd try a 3B model, store meals in a small file so memory doesn't depend on chat length, and test it with a real classmate.
Once my PC is fixed, I'd like to turn this into a proper build with a bigger model and a saved meal log, so the memory doesn't depend on a chat staying short.
Screenshots
(Attach: the cilantro loop, the Tuesday entry reply, and the Friday reply.)
Top comments (0)