Telegram bot for people with diabetes: send a food photo, get carbs, calories, portion estimate, and an insulin bolus suggestion.

FoodShot is a practical Telegram assistant for meal estimation. A user sends a photo of food, the bot recognizes the dish, estimates the portion, fetches nutrition data, asks for current glucose if needed, and calculates a suggested insulin dose using the user's personal settings.

FoodShot gives estimates only. It is not a medical device and does not replace advice from a qualified healthcare professional.

The bot works directly inside Telegram:

• recognizes the dish from a photo;
• estimates portion weight and confidence;
• calculates carbs and calories;
• lets the user adjust the portion with inline buttons;
• asks for current blood glucose or accepts /skip;
• returns a bolus breakdown and saves the meal history.

Photo in Telegram
  -> FastAPI webhook receives Telegram update
  -> aiogram handler downloads the image
  -> Vision AI detects dish, portion, confidence
  -> Nutrition service fetches macros
  -> User adjusts portion and enters glucose
  -> Calc engine computes bolus
  -> Result is sent back and saved to PostgreSQL

FoodShot keeps the calculation transparent:

carb_dose  = carbs_g / icr
correction = (current_bg - target_bg) / isf
total      = carb_dose + correction

Correction is used only when the user provides current glucose and it is above the target.

.
├── api/
│   └── webhook.py          # FastAPI app and Telegram webhook
├── bot/
│   ├── handlers/           # start, photo, settings, history, help
│   ├── keyboards/          # reply and inline keyboards
│   ├── states.py           # aiogram FSM states
│   └── i18n_middleware.py  # language detection
├── core/
│   ├── config.py           # pydantic settings
│   └── i18n.py             # runtime translations
├── db/
│   ├── models.py           # SQLAlchemy models
│   ├── crud.py             # async DB operations
│   └── database.py         # engine and session setup
├── services/
│   ├── vision.py           # OpenAI vision wrapper
│   ├── nutrition.py        # nutrition data client
│   └── calc.py             # bolus calculation
├── doc/
│   ├── foodshot_architecture.svg
│   └── tg_fries.png
├── docker-compose.yml
├── Dockerfile
├── Taskfile.yml
└── pyproject.toml

Create .env from the example:

cp .env.example .env

Required values:

BOT_TOKEN=your_telegram_bot_token
OPENAI_API_KEY=your_openai_api_key
USDA_API_KEY=your_usda_api_key

POSTGRES_USER=foodshot
POSTGRES_PASSWORD=foodshot_pass
POSTGRES_DB=foodshot

DATABASE_URL=postgresql+asyncpg://foodshot:foodshot_pass@localhost:5432/foodshot
REDIS_URL=redis://localhost:6379/0

WEBHOOK_URL=https://your-ngrok-domain.ngrok-free.app/webhook
PORT=8000

For Docker Compose, the bot container overrides database and Redis hosts internally:

DATABASE_URL -> db:5432
REDIS_URL    -> redis:6379

That keeps .env convenient for local development while Docker services still talk to each other by service name.

Install dependencies:

task install

Start PostgreSQL and Redis only:

task infra

Run FastAPI locally with reload:

task dev

Run the full Docker stack:

task up

Show logs:

task logs

Stop services:

task down

Expose port 8000 with ngrok:

ngrok http 8000

Set the Telegram webhook URL in .env:

WEBHOOK_URL=https://your-ngrok-domain.ngrok-free.app/webhook

The app registers this URL on startup:

POST /webhook

If Telegram keeps calling an old URL, inspect it with:

curl -s "https://api.telegram.org/bot$BOT_TOKEN/getWebhookInfo"

FoodShot stores:

• users: Telegram id, username, ICR, ISF, target glucose, insulin type, language.
• meal_logs: dish, portion, macros, suggested bolus, current glucose, photo id, timestamp.

Food recognition, portion estimation, nutrition databases, and insulin formulas can be wrong. FoodShot is an assistant for estimation, not a clinical decision system.

Always verify calculations before making insulin decisions and follow the treatment plan agreed with a qualified healthcare professional.