My roommate Nagaraj needed high-protein vegetarian food on a ₹2000/week budget — while strictly avoiding peanuts and lactose. So for the Build for a Friend theme, I built him a meal planner that runs entirely on his laptop: no accounts, no API keys, no cloud.
Friend Meal Planner — a web app (vanilla HTML/JS + a zero-dependency Python server, stdlib only) that generates a 7-day Indian meal plan with a grocery list and budget check. You fill in your friend's profile — diet, allergies, dislikes, budget — and the plan streams in, token by token.
python3 app.py and open http://localhost:8000 (needs Ollama + gemma3:1b pulled; ~40s per plan on CPU)
Nagaraj, my roommate in Bengaluru. Hostel-style cooking, 30-minute recipes, ingredients from local markets, everything priced in ₹. I handed him the first plan — his verdict: "now that's cooking" (he also immediately asked for a recipe mode and non-veg mode, so that's next).
Open-weight AI is what makes this project possible, not just cheaper:
Testing caught something important: the 1B model kept suggesting paneer and yogurt to my lactose-allergic friend, despite explicit instructions. Rather than hide that, I added a deterministic allergy guardrail — every plan is scanned line-by-line against an allergen keyword map (lactose → milk, curd, paneer, ghee…) and an ⚠️ Allergy Check section is appended flagging risky dishes for review. Prompt engineering sets the intent; code enforces the safety property. That's the architecture I'd defend: never let a probabilistic model be the last word on food safety.
This also enters the Best Use of Gemma category — the whole app is Gemma-powered, running the open-weight gemma3:1b model locally through Ollama.