My Projects

Here is an non exhaustive list of projects I've worked on.

RobotHub

RobotHub is an open-source platform that brings real-time robot control, AI inference and collaborative 3D visualisation directly to your browser.

Built with SvelteKit, Threlte and WebRTC, RobotHub lets you connect USB robots, stream cameras, and run AI policies – all through a shareable URL like #workspace-abc. It powers live demos hosted on Hugging Face Spaces and runs 100 % in the browser.

2025 Robotics WebRTC SvelteKit
RobotHub 3D interface showing robot and status HUDs
Fig. 1: Live multi-robot control in the browser

Arxflix

Arxflix is my AI Paper Reading Youtube Channel.

Born at a Paris hackathon in May 2024, ArxFlix is a fully automated video-generation agent that transforms academic papers into engaging, two-minute summaries. It parses complex text, figures, and equations, scripts a clear narrative, and delivers it all with a smooth, human-like voiceover—making videos.

2024 Hackathon Mixtral-8x7B LoRA Text-to-Speech React (Remotion)
Fig. 1: Example of arxflix video

Montelimar

Montelimar: an open-source, on-device OCR toolkit for snipping and copying text—flexible, modular, and perfect for non-Latin scripts and LaTeX.

Montelimar is an open-source, on-device OCR engine and desktop application that lets you snip a portion of your screen and instantly copy the recognized text to your clipboard. It's designed as a general-purpose, unopinionated toolkit with a flexible, modular architecture—ideal for anyone building custom OCR workflows. My goal is for Montelimar to become the go-to framework for next-generation OCR apps, especially those targeting non-Latin scripts and LaTeX-formatted content.

2024 Nougat OCR Application Svelte Rust MLX
Fig. 2: Montelimar interface

PixDiet

A finetune of Pixtral 12B VLM on dietetic data to generate a personalized meal plan based on your food preferences and health goals.

Again a 24h Hackathon project. I finetuned a Pixtral 12B model on dietetic data to generate a personalized meal plan based on your food preferences and health goals. It was the first finetune of Pixtral model just 24 hours after the open released.

2024 Hackathon Pixtral 12B LoRA
PixDiet UI showing generation results
Fig. 3: Pixtral before and after finetuning on dietetic data

Furniture Inpainting

Seamlessly blend furniture into images for e-commerce and interior design.

Custom train a Flux adapter to inpaint a reference furniture (from an image) into a scene. The adapter handle by itself the furniture orientation, the lighting and the perspective.

2025 DiT ControlNet Modal Quantization
Montelimar app interface with recognized text
Fig. 4: Automatic furniture inpainting with lighting and perspective adaptation

Virtual Staging

Transform empty rooms into fully furnished spaces—no masking required!

A mask-free solution that transforms empty rooms into fully furnished spaces—revolutionizing how real estate professionals showcase properties. The adapter is designed to preserve the original scene as much as possible. By fusing the entire virtual staging pipeline into a single adapter, we can process an image in under 10 seconds.

2025 DiT ControlNet Modal Quantization
Room virtually staged with furniture
Fig. 5: Empty room transformation with automatic furniture placement

Image Decluttering

Remove unwanted objects from interior photos in seconds.

A mask-free solution, similar to the virtual staging, but for decluttering.

2025 DiT ControlNet Modal Quantization
Decluttered room after object removal
Fig. 6: Automatic removal of unwanted objects from interior scenes

Socials

Links

Miscellaneous

  1. [1] All opinions are my own, except those generated by large language models.
  2. [2] Fonts: ...
Guybrush.ink
Made with ♥ in Paris, London & Toulouse build: main