Hafedh Hichri

Hafedh Hichri, also known online as not-lain, is a Software Engineer at Chonkie and a Hugging Face Fellow. He develops custom AI models and shares his knowledge through blog posts on topics like RAG and custom model training.

He is passionate about making complex research accessible through open-source software, contributing to libraries like transformers, unsloth, and autotiktokenizers. He studied Computer Science at the National School of Electronics and Telecommunications of Sfax (ENET'Com).

Blogs

I write about machine learning, natural language processing, and open-source software mostly in HuggingFace. Here are some of my recent blog posts:

• 👁️ Visualizing How VLMs Work A deep dive into how VLMs aggregate and process data across modalities, co-authored with Ed Daniels.
• 🔭 KV Caching Explained: Optimizing Transformer Inference Efficiency A deep dive into the concept of KV caching in transformers, explaining its significance and providing practical examples.
• 🔍 Mastering Tensor Dimensions in Transformers A comprehensive guide to understanding tensor dimensions in transformers, with practical examples and tips for effective manipulation.
• 🚀 PyTorchModelHubMixin: Bridging the Gap for Custom AI Models on Hugging Face A detailed exploration of the PyTorchModelHubMixin class, showcasing its role in integrating custom AI models with the Hugging Face ecosystem.
• 🧠 RAG using huggingface tools A step-by-step guide to building a Retrieval-Augmented Generation (RAG) chatbot using Hugging Face tools, with practical examples and code snippets.
• 🚀 Image-based search engine A tutorial on creating an image-based search engine using Hugging Face tools, with practical examples and code snippets.

Open Source

Most of my work comes in this field the form of contributions to other libraries. A couple of notable libraries I contributed to are:

🤗 Transformers

State-of-the-art AI library by Hugging Face.

🦥 Unsloth

Efficient fine-tuning for LLMs.

Gradio

Python Library for building machine learning web-applications.

Chonkie [maintainer]

AI library for efficient data chunking and building RAG pipelines.

Note: Currently working at this company and serving as their maintainer.

🤗 HuggingFace.js

JavaScript client for Hugging Face APIs. I contributed to API features and bug fixes.

fal

SDK Client for FAL

I also maintain a couple of libraries:

Loadimg

A fast, lightweight image loader for web apps. I am the creator and maintainer.

Pxia

A pixel manipulation library for creative coding. I am the creator and maintainer.

I also contributed to integrating several models with Hugging Face using PyTorchModelHubMixin, including:

BiRefNet

SOTA Background Removal model for camouflaged image segmentation (chameleon, etc.).

BEN2

SOTA Background Removal model for detailed image segmentation (hair strands, etc.).

MatAnyone

Video propagation model for consistent object segmentation across frames.

anime-seg

Anime image segmentation model.

araclip

Arabic CLIP model for image-text tasks.

RMBG-1.4

Background removal model.

SwarmFormer

Transformer variant using hierarchical local-global attention reducing cost with strong accuracy.

Invited Talks and News

• On October 16, 2025, I gave a speech at PyData about AI agents [slides]
• On October 9th, 2025, I gave a talk at the Hugging Face server about Building how VLMs work [slides].
• On June 13 to 15 2025, I was one of the judges at Artificial Intelligence National Summit V3 and I gave a talk about Chonkie [slides].
• On April 27, 2025, I gave a talk at the Genesis Labs-INSAT about AI in Healthcare [slides].