Two weeks ago, I launched my first app Nutix Ai, a calorie tracking app focused on simplicity. It's now at 200 downloads, which feels great for a first launch.
Today I want to share one specific part of the journey: building a food database that supports multiple languages (English, Arabic, Spanish) — not just English like most nutrition apps.
Why Multilingual?
Most calorie tracking apps have English-only databases, or their translations feel like afterthoughts. I wanted native speakers to actually be able to search for foods in their language from day one.
The Architecture
Simple setup:
- Local LLM running on my AMD 9070XT (yes, not NVIDIA — and it worked great)
- Base food data in English
- LLM translates each entry to Arabic and Spanish
- Output stored in my database
The Model Hunt (The Painful Part)
Finding the right model took longer than I expected.
| Model | Speed | Translation Quality (especially Arabic) |
|---|---|---|
| DeepSeek 7B | Slow | Terrible |
| DeepSeek 8B | Slow | Terrible |
| DeepSeek 14B | Very slow | Still bad |
| Llama variants | Moderate | Poor |
| QWEN | Had high hopes | Performed worse than expected |
| Gemma3 7B | ~200 tok/s | Good (1-5% error rate) |
| Gemma3 12B | ~80 tok/s | Great |
Gemma3 was the winner. The 7B was fast but had occasional errors. The 12B took about 2 seconds per food item but was much more reliable.
The Result
~6,000 foods translated into 3 languages in 4 hours, all generated locally on consumer hardware.
Known Issues & What's Next
- Database size — 6K foods is a start, but I want to expand to 12+ languages and more regional foods
- Search — The food search still needs work. Fuzzy matching across multiple languages is tricky
Looking for Suggestions
If you have ideas for:
- Good sources of food/nutrition data
- Improving multilingual search
- Other languages I should prioritize
Drop them in the comments — I'd genuinely appreciate it.
You can check out the app here: Nutix on the App Store
Top comments (0)