MyMe — La Comète du Mbam

A custom single-restaurant web app: paste today's menu (like a WhatsApp message), click one button, get beautiful PNG posters ready for WhatsApp Status / Instagram / Facebook.

Every submission generates five posters — four hand-crafted templates (Organic, Elegant, Bold, Broadsheet) plus one AI-designed poster created on the fly by the LLM itself using your brand identity tokens. Download whichever looks best that day.

---

How it works

Staff pastes the menu text
        ↓
GLM-5.2 (via NVIDIA NIM + OpenAI SDK) turns it into structured JSON
   { title, mains[], complements[], drinks[] }
   (LLM also picks the best matching photo filename per dish)
        ↓
Each dish is matched to a photo in media/foods/ by filename.
If a hero dish has no local photo, the image search waterfall runs:
   Google Custom Search → DuckDuckGo → Unsplash
   The first hit is downloaded and permanently cached in media/foods/.
        ↓
Five posters are rendered:
  • organic    — white bg, leaf decorations, circular dish photos
  • elegant    — cream bg, gold rules, hero photo strip
  • bold       — high-contrast, large type
  • broadsheet — newspaper-style layout
  • dynamic    — designed from scratch by the LLM using brand tokens ✦
        ↓
Playwright (headless Chromium) screenshots each to a 1080×1350 PNG
        ↓
Staff downloads the PNG they prefer

---

1. Setup

python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

pip install -r requirements.txt
python manage.py setup_browser  # installs Chromium for Playwright (one-time)

cp .env.example .env
# Open .env and add your NVIDIA_API_KEY
# (get one free at https://build.nvidia.com/)

python manage.py migrate
python manage.py runserver

Open http://127.0.0.1:8000 — paste a menu, click "Générer l'affiche", download.

---

2. Environment variables

Variable	Required	Description
NVIDIA_API_KEY	Yes	NVIDIA NIM key for GLM-5.2 inference
DJANGO_SECRET_KEY	Yes (prod)	Long random string for Django sessions
GOOGLE_SEARCH_API_KEY	Optional	Upgrades image search to Google CSE
GOOGLE_SEARCH_ENGINE_ID	Optional	Goes with the key above
RENDER_BASE_URL	Optional	Playwright base URL (default: http://127.0.0.1:8000)

---

3. The LLM stack

• SDK: OpenAI Python SDK — same interface works with OpenAI, Anthropic, Mistral, NVIDIA NIM, and any other OpenAI-compatible endpoint.
• Model: zhipuai/glm-4-9b-chat (GLM-5.2) hosted on NVIDIA NIM.
• Switch models: change LLM_MODEL + NVIDIA_BASE_URL + NVIDIA_API_KEY in config/settings.py.

---

4. Image search waterfall

When a dish has no local photo and is needed as a hero, three sources are tried:

1. Google Custom Search (optional) — highest quality, needs API key
2. DuckDuckGo images (free, no key) — solid results for most dishes
3. Unsplash Source (free, no key) — generic food photo as last resort

All downloads are cached permanently in media/foods/ so future menus never re-search the same dish.

---

5. Dynamic AI poster

The LLM receives layout skills (rules for great 1080x1350 menu poster structure), brand identity tokens (palette, typography, mood, hard constraints), and the full menu data with image URLs. It returns a complete HTML/CSS document which is Django-template-rendered server-side then screenshotted by Playwright.

The brand identity system lives in BRAND_IDENTITY in config/settings.py. Adapt it per restaurant by swapping colour values and fonts.

---

6. Project structure

config/
  settings.py          Django settings + BRAND_IDENTITY + LLM config

menu/
  llm_service.py       OpenAI-SDK client for GLM-5.2 via NVIDIA NIM
  image_matcher.py     Filename-based dish to photo matching
  image_search.py      Image search waterfall: Google, DuckDuckGo, Unsplash
  poster_renderer.py   HTML to PNG via Playwright (static + dynamic modes)
  models.py            GeneratedMenu history record
  views.py             home / generate_menu / download_poster
  templates/menu/
    posters/           organic, elegant, bold, broadsheet (hand-crafted)
  static/menu/
    img/               logo.png, default-dish.png

media/
  foods/               Real dish photos (matched by filename)
  generated/           Output PNGs land here