1 of 6

PixDiet

A personalised AI nutritionist in your pocket

🥕

🍓

🥦

2 of 6

2

WHO: Everybody needs nutrition advices

From people with chronic disease (obesity, renal failure…) to startupers running triathlon everybody can use Nutritional advices… Yet less than half of French people consult a dietician over their lifetime..

Matthieu

  • Age: 27yo
  • Weight: 75kg

Medical condition:

  • Stage 3 chronic kidney disease

Objectives:

  • Maintain good health

Persona 1: Chronic disease patient

Julien

  • Age: 24yo
  • Weight: 70kg

Medical condition:

  • Workalcoholic

Objectives:

  • Maintain maximum energy levels during hackathon and minimize sleep.

Persona 2: Hard working startuper

Carla

  • Age: 52yo
  • Weight: 58kg

Medical condition:

  • Overweight

Objectives:

  • Training for her first marathon

Persona 3: first time marathon runner

🧑‍🦰

🧑‍🦱

🧔‍♀️

💵

“it's too expensive”

“I am not really sick/ legitimate”

“Pre-made programs are too constraining”

3 of 6

Introducing PixDiet a Personalised AI nutritionist in your pocket

Realtime: replies directly

Multimodal: understands images and can classify food

Personalised: knows your meal history and med condition

Domain expert: trained on nutritionist expert knowledge.

4 of 6

How we built it ?

Pixtral 12b

The first ever finetune of Pixtral 12b !

PixDiet

food pic

items

dataset: 6k pairs of (image, items)

AI generated nutritionist comment

pasta, cheese

….

….

this pasta salad provides a good mix of carbs from the fusilli, fats from the cheese, and some vitamins from the rocket and sun-dried tomatoes. However, as a marathon runner, you might need more protein and fiber to support muscle recovery and sustained energy ……….

  • Synthetic dataset: 6000 samples
  • Parameters: Training on 1% of parameters (LoRa)
  • Epoch: 3
  • Hardware: NVIDIA H100 80GB (thanks nebius)

Step 2: Training

Step 1: Synthetic database creation

  • Transformed tabular dataset to Instruct dataset
  • Infused expert knowledge and tone to align the model with the role of a nutritionist
  • Open source live demo on huggingface
  • Fully running on GPUs zero

Step 3: Demo

🥕

🍓

🥦

Synthetic dataset creation, finetuning, gradio demo running fully on HuggingFace space

5 of 6

5

Appendix: Before/ After finetune (qualitative benchmark)

6 of 6

6

Appendix (links)