Website Home page
Website Home page
Whatsapp Bot
Whatsapp Bot
Whatsapp Bot
Whatsapp Bot meal recommendation
Whatsapp Bot meal recommendation

Foodie Robot

Inspiration

Food ordering apps hand you a menu and leave you guessing — will this fit my fitness goals? is this too many calories? Foodie Robot flips that around: it recommends meals tailored to your fitness goals and lets you order them directly, so it eliminates the guesswork.

And it lives where you already are — WhatsApp. No new app, no new account. You just text it like a friend who knows every restaurant in town — "something light and high-protein for around ₦3000" — and it tells you exactly what to eat and lets you order it on the spot.

What it does

Chat with the bot on WhatsApp in plain language.
It recommends meals tailored to your fitness goals, taste, budget, and time of day — and you order them directly, no guesswork.
It can analyze a meal photo to estimate calories and nutrition.
It handles the full flow — location, ordering, payments, referrals, and balances.

How I built it

Backend: Django (ASGI / Daphne) with a Huey worker for background jobs, PostgreSQL + PostGIS for data and geo, and Redis for caching and the task queue.
AI: Alibaba Cloud's Qwen models via Model Studio (DashScope), used in three ways:
- qwen-max for chat and tool calling — the bot decides which action to run (recommend, order, save location, etc.).
- qwen-vl-max for vision — reading a meal photo.
- text-embedding-v4 for embeddings — turning meals and requests into vectors so I can match them by meaning.
Hosting: Docker containers on Alibaba Cloud (ECS), with managed RDS (PostgreSQL) and Redis.

Challenges I ran into

Structured output from vision. Getting reliable, schema-shaped JSON out of the image analysis took prompt tuning plus a forgiving parser (handling code fences and stray text) so a single odd response wouldn't break the flow.
Embedding dimensions. My database assumed 1536-dim vectors, so I pinned the embedding size to keep everything consistent.
Server-side image fetching. The vision model downloads image URLs itself, so images must be hosted somewhere reliably reachable — I use a global CDN (Cloudinary) for that.
Keeping costs low. Caching embeddings and filtering tools before each call kept token usage (and the bill) small.

Accomplishments that I'm proud of

A bot that feels natural. Tool calling lets it understand free-form requests and take real actions — no rigid menus or button trees.
Recommendations that actually fit you. Semantic matching on fitness goals, taste, and budget turns "what should I eat?" into a direct answer you can order.
Full end-to-end flow on WhatsApp. From recommendation to location to payment to referrals — all inside a chat, with no extra app to install.
Lean and cost-efficient. Cached embeddings and pre-call tool filtering keep it fast and cheap to run.

What I learned

Tool calling beats menus. Letting the model pick the right action from a set of well-described tools made the bot feel natural instead of scripted.
Embeddings make search feel smart. A little linear algebra goes a long way — semantic matching is far better than keyword matching for food.
Qwen is genuinely capable across chat, vision, and embeddings, and one API key covers all three.