AI Fablab & Demonstrator

INSA Rennes Logo Teklia Logo



An innovative educational project to explore, understand and test artificial intelligence freely.

5 AI Tasks
10 AI Models
4 Themes

About the project

The AI Fablab project was born from a simple observation: AI technologies are increasingly numerous and complex. Our mission is to make these tools accessible to everyone, from beginners to experts.

For this first year, the AI Fablab focuses on image processing.

Our mission

Democratize access to artificial intelligence tools by creating a learning space where everyone can discover, understand and experiment with different AI models.

Our vision

Become the INSA reference for hands-on learning of applied AI across multiple domains, offering a unique practical learning experience.

Academic context

This project is part of the INSA Rennes 4INFO curriculum, in partnership with Teklia, a company specializing in document image analysis using artificial intelligence.

Duration: 2 semesters (September 2025 - May 2026)

2025-2026 theme: Image processing (future extensions: audio, video, text)

The AI Fablab

An interactive and educational encyclopedia dedicated to artificial intelligence models for image processing.

What is the AI Fablab?

The AI Fablab is a web platform that centralizes and documents AI models. Inspired by physical FabLabs that provide tools and knowledge, our virtual Fablab guides you from theory to practice.

Organization by domains

The AI Fablab separates tasks and models into different categories, with a modular architecture ready to support additional domains (audio, video, text).

Documented tasks

5 classical tasks: object detection, OCR, segmentation, image similarity, and more.

Referenced models

A curated selection of models (YOLO, CLIP, ResNet, TesseractOCR...) is available in the AI Fablab.

Covered themes

Theme 1: Image and text processing

Image analysis via keywords (CLIP, SigLIP), OCR (Tesseract, Paddle OCR, Qwen2.5-VL)

Theme 2: Object and face recognition

Classification (ResNet-50), face detection, depth estimation (Depth Pro), pose estimation (VITPose)

Theme 3: Image similarity

Image embeddings (ViT, DINOv3), similarity metrics (cosine, SSIM)

Theme 4: Image quality analysis

Aesthetic evaluation using HumanAesExpert and MUSIQ

Model documentation content

Each model is documented with maximum detail:

Administration interface

The AI Fablab includes a secure admin interface to easily manage all content (domains, tasks, models).

It allows creating, editing and deleting entries directly from the browser, without technical skills or external tools.

Explore the AI Fablab

Model input/output details

Model input/output details

Model resources links

Model resources links

Fablab admin page

Fablab admin page

The Demonstrator

An interactive application to test AI models on a collection of INSA Rennes photo archives.

What is the Demonstrator?

The demonstrator is a web application that applies models from the Fablab in practice. It lets users explore real image archives (INSA 60-year archives) using AI tasks.

Learning by doing philosophy: theoretical concepts from the Fablab become directly testable in the demonstrator.

Three demonstration levels

Level 1: Simple search

Image search engine using keywords. AI is transparent to the user, like Google Images.

Technologies: CLIP, SigLIP encoders, precomputed AI filters

Level 2: Comparison & chaining

Selection of tasks and models with ability to:

  • compare multiple models,
  • adjust parameters,
  • chain processing tasks.

Level 3: Fablab redirection

Detailed explanation of how the demonstrator is built.

Direct links to the Fablab for deeper understanding and full documentation.

Performance optimization

Issue: running AI models in real time is too expensive for multiple users.

Solution: heavy precomputation

Result: near-instant responses even for complex queries (on ~4500 images dataset).

Explore the Demonstrator

Simple search (level 1)

Simple search page

Advanced search (level 2)

Advanced search page

Expert mode (level 3)

Expert search page

Technical architecture

A modular and scalable architecture designed for maintainability and extensibility.

Fablab / Demonstrator separation

Two independent web applications:

  • Fablab: static documentation + lightweight backend (FastAPI)
  • Demonstrator: interactive app with optimized backend

Each has its own frontend (Angular) and backend (Python/FastAPI).

Shared library

Common Python module that:

  • abstracts AI model usage
  • unifies input/output interfaces
  • avoids code duplication
  • simplifies adding new models

Data storage

Fablab: Hybrid

  • SQL database: metadata and relationships
  • JSON files: educational content (flexible structure and external links)

Demonstrator: Optimized

  • SQLite: images, precomputations (read-only)
  • .npy files: embeddings (NumPy binary format)

Containerized deployment

Docker for:

Hosting: VPS provided by INSA Rennes IT department

Technologies used

Frontend

Angular
TypeScript framework

TypeScript
Static typing

HTML5 / CSS3
Structure & design

Backend

Python
Main language

FastAPI
REST API framework

SQLite
Database

NumPy
Numerical computing

Artificial Intelligence

PyTorch
Deep Learning

Transformers
Hugging Face

OpenCV
Computer Vision

Tools & DevOps

Docker
Containerization

Git / GitLab
Version control

VSCode
Code editor

Team & Partners

Development team

INSA Rennes supervisors

INSA Logo
Eric Anquetil Academic supervisor - Computer Science Department
INSA Logo
Christian Raymond Academic supervisor - Computer Science Department

Industrial partner: Teklia

Teklia is a company specialized in automatic document analysis using artificial intelligence.

Contributions:

  • Expertise in document image processing
  • Feedback on AI models
  • Architecture and best practices guidance
  • Access to HikarIA project (inspiration)
Teklia Logo
Mélodie Boillet Teklia supervisor - AI engineer
Teklia Logo
Solène Tarride Teklia supervisor - AI engineer