AI-Assisted Report

Rapport assiste par IA

Creation de documents avec des agents IA

Creating Documents with AI Agents

Denis Henin

Accenture Song

~30 min

Sommaire

Overview

Menu de la presentation

Presentation menu

Claude Instructions

Framework d'agents IA

AI agent framework

Architecture

Structure du projet

Project structure

Thesaurus & Mindmap

Concepts et termes

Concepts and terms

Sources de Recherche

Research Sources

Bases de donnees academiques

Academic databases

Analyse Multi-Agents

Multi-Agent Analysis

5 agents pour analyse PDF

5 agents for PDF analysis

Tech Stack

LaTeX, Python, Ollama

Resultat Final

Final Result

34 pages, v2.6

Autres Projets

Other Projects

News, JurisAssist, Equi

Comment le rapport est produit

How the report is produced

Pipeline en 5 etapes

5-step pipeline

Recherche

Research

40+ PDFs

Analyse IA

AI Analysis

5 agents

LaTeX

Redaction

Writing

Design HTML

HTML Design

Playwright

PDF Final

Final PDF

v2.6.0

Claude Instructions

Framework pour travailler efficacement avec Claude Code

Framework for working efficiently with Claude Code

Agents specialises

Specialized agents

Architecture hierarchique

Hierarchical architecture

Claude Instructions

Framework d'agents pour Claude Code

Agent Framework for Claude Code

Claude Instructions

Framework d'agents

Agent framework

Reutilisabilite maximale, qualite constante, montee en competences acceleree des agents.

Architecture hierarchique: Les agents heritent des regles globales (securite OWASP, code style) definies dans /rules, puis chargent leurs propres configurations YAML.

Agents specialises: Senior Fullstack (W3C, WCAG), UX Expert (Playwright, Front-End Checklist), Memory Manager (LightRAG, knowledge graphs).

Integration MCP: Connexion aux outils externes - bases de donnees, navigateurs headless, APIs tierces - via le protocole Model Context.

Frameworks inclus: AI-Manager (orchestration Python), Better-Agents (CLI TypeScript), liens vers Google ADK.

Maximum reusability, consistent quality, accelerated agent skill development.

Hierarchical architecture: Agents inherit global rules (OWASP security, code style) defined in /rules, then load their own YAML configurations.

Specialized agents: Senior Fullstack (W3C, WCAG), UX Expert (Playwright, Front-End Checklist), Memory Manager (LightRAG, knowledge graphs).

MCP integration: Connection to external tools - databases, headless browsers, third-party APIs - via Model Context Protocol.

Included frameworks: AI-Manager (Python orchestration), Better-Agents (TypeScript CLI), links to Google ADK.

09-CLAUDE-INSTRUCTIONS/
├── agents/
│   ├── general-work/
│   │   ├── CLAUDE.md, AGENTS.md
│   │   ├── prompts/ (yaml, json)
│   │   └── docs/ (workflow, debug, git)
│   ├── web-development/
│   │   ├── senior-fullstack/ (W3C, WCAG, OWASP)
│   │   ├── architecture/ (patterns, hexagonal)
│   │   └── accessibility/ (ARIA, testing)
│   ├── ux-expert/ (Playwright, checklists)
│   └── memory-manager/ (LightRAG, FalkorDB)
├── frameworks/
│   ├── ai-manager/ (Python orchestrator)
│   ├── better-agents/ (TypeScript CLI)
│   └── google-adk-*/ (symlinks vers Agent Dev Kit)
├── rules/
│   ├── claude-rules.md
│   ├── code-style.md
│   └── security.md
└── docs/ (INDEX, MIGRATION-GUIDE)
        

Architecture

Structure du projet et organisation des fichiers

Project structure and file organization

Methodologie_de_recherche/ ├── 01-portfolio/ ├── 02-design-pages/ ├── 03-presentation/ ├── design-system/ ├── research/ │ └── agents/ └── references/

Structure du Projet

Project Structure

Methodologie_de_recherche/
├── 01-portfolio/
│   ├── latex/src/           # Sources LaTeX
│   └── releases/            # 20+ versions
├── 02-design-pages/
│   ├── sources/             # Pages HTML
│   └── export-pdf.js        # Playwright
├── 03-presentation/         # Reveal.js
├── course_materials/        # MDRD
├── design-system/           # Design tokens
├── references/              # Bibliographie
└── research/
    ├── agents/              # 5 agents IA
    └── sources/             # 40+ PDFs
        

Methodologie_de_recherche/
├── 01-portfolio/
│   ├── latex/src/           # LaTeX sources
│   └── releases/            # 20+ versions
├── 02-design-pages/
│   ├── sources/             # HTML pages
│   └── export-pdf.js        # Playwright
├── 03-presentation/         # Reveal.js
├── course_materials/        # MDRD
├── design-system/           # Design tokens
├── references/              # Bibliography
└── research/
    ├── agents/              # 5 AI agents
    └── sources/             # 40+ PDFs
        

Recherche

Research

Sources academiques et analyse IA

Academic sources and AI analysis

Design

Pages visuelles HTML/CSS

Visual HTML/CSS pages

Production

Compilation LaTeX finale

Final LaTeX compilation

Design System

Architecture CSS Myerscough

Myerscough CSS Architecture

design-system/

main.css
├── _variables.css
├── _typography.css
├── _components.css
├── _patterns.css
├── _utilities.css
└── README.md
          

Or#FFE500

Cramoisi#FF0050

Azur#0066FF

Orange#FF6600

Violet#FF00AA

Lime#00FF66

Gold#FFE500

Crimson#FF0050

Azure#0066FF

Orange#FF6600

Violet#FF00AA

Lime#00FF66

Design Vibrant

Vibrant Design

Couleurs Myerscough

Myerscough colors

Rigueur LaTeX

LaTeX Rigor

Typo pro, APA

Pro typography, APA

Illustrations

Mindmap, thesaurus

Thesaurus & Mindmap

Premiere etape : definir les concepts et termes

First step: defining concepts and terms

Construction du Thesaurus

Thesaurus Construction

Methodologie AFNOR - Avant l'analyse PDF

AFNOR Methodology - Before PDF analysis

Definition du sujet

Subject definition

Iconographie des ailes dans la peinture

Wing iconography in paintings

Collecte des termes

Term collection

Brainstorming, recherche preliminaire

Brainstorming, preliminary research

Organisation hierarchique

Hierarchical organization

Relations entre termes (generique/specifique)

Term relationships (generic/specific)

Visualisation Mindmap

Mindmap Visualization

Mind Elixir + theming custom

Mind Elixir + custom theming

Mindmap Interactive

Interactive Mindmap

Generee avec Claude Code

Generated with Claude Code

140+ termes organises hierarchiquement
Librairie Mind Elixir (JavaScript)
Export PNG via Playwright
Integree dans le PDF final

140+ terms organized hierarchically
Mind Elixir library (JavaScript)
PNG export via Playwright
Integrated into final PDF

Sources de Recherche

Research Sources

Collecte des PDFs academiques

Collecting academic PDFs

40+

Sources PDF

PDF Sources

Bases de donnees

Databases

Pipeline de Recherche

Research Pipeline

Agents IA & Sources

AI Agents & Sources

research/

├── agents/
│   ├── analyseur_pdf_auto.py
│   └── prompts/
│       ├── AGENT-0-ORCHESTRATEUR.md
│       ├── AGENT-1-TRIEUR.md
│       ├── AGENT-2-EXTRACTEUR.md
│       ├── AGENT-3-ICONOGRAPHIQUE.md
│       └── AGENT-4-TEXTE.md
├── sources/
│   ├── a_traiter/           # A analyser
│   ├── analysees/           # Completees
│   └── hors_sujet/          # Rejetees
└── docs/
          

├── agents/
│   ├── analyseur_pdf_auto.py
│   └── prompts/
│       ├── AGENT-0-ORCHESTRATEUR.md
│       ├── AGENT-1-TRIEUR.md
│       ├── AGENT-2-EXTRACTEUR.md
│       ├── AGENT-3-ICONOGRAPHIQUE.md
│       └── AGENT-4-TEXTE.md
├── sources/
│   ├── a_traiter/           # To analyze
│   ├── analysees/           # Completed
│   └── hors_sujet/          # Rejected
└── docs/
          

00 ORCHESTRATEUR ORCHESTRATOR

Coordination & Synthese

Coordination & Synthesis

01 TRIEUR SORTER

Evaluation MDRD

MDRD Evaluation

02 EXTRACTEUR EXTRACTOR

Themes & mots-cles

Themes & keywords

03 ICONOGRAPHIQUE

Analyse Panofsky

Panofsky Analysis

04 TEXTE TEXT

Structuration MD

MD Structuring

Analyse PDF Multi-Agents

Multi-Agent PDF Analysis

5 agents pour analyser les sources collectees

5 agents to analyze collected sources

Agents IA au total

AI Agents Total

Fichiers MD par PDF

MD files per PDF

40+

Sources analysees

Sources analyzed

Pipeline d'Analyse PDF

PDF Analysis Pipeline

Source PDF

PDF Source

Agent 1

Agent 2

Agent 3

Agent 4

5 Fichiers MD

5 MD Files

AGENT 1 TRIEUR SORTER

Evaluation MDRD : Autorite, Actualite, Objectivite, Exactitude

MDRD Evaluation: Authority, Currency, Objectivity, Accuracy

01-evaluation-MDRD.md

AGENT 2 EXTRACTEUR EXTRACTOR

Extraction thematique : oeuvres, artistes, periodes, mots-cles

Thematic extraction: artworks, artists, periods, keywords

02-extraction-thematique.md

AGENT 3 ICONOGRAPHIQUE

Analyse des oeuvres selon grille 5 etapes (Panofsky)

Artwork analysis using 5-step grid (Panofsky)

03-analyse-iconographique.md

AGENT 4 TEXTE TEXT

Structuration Markdown, numerotation, indexation

Markdown structuring, paragraph numbering, indexing

04-analyse-texte.md

Implementation Technique

Technical Implementation

# analyseur_pdf_auto.py

def process_pdf(pdf_path):
    # 1. Extraction texte (PyPDF2)
    text = extract_text_from_pdf(pdf_path)

    # 2. Appels sequentiels des agents
    response_1 = agent_1_trieur(text)
    response_2 = agent_2_extracteur(text)
    response_3 = agent_3_iconographique(text)
    response_4 = agent_4_texte(text)

    # 3. Sauvegarde des resultats
    save_markdown_files(responses)

# Modele : Ollama + qwen3:32b
# Timeout : 30 minutes par agent
        

# analyseur_pdf_auto.py

def process_pdf(pdf_path):
    # 1. Text extraction (PyPDF2)
    text = extract_text_from_pdf(pdf_path)

    # 2. Sequential agent calls
    response_1 = agent_1_trieur(text)
    response_2 = agent_2_extracteur(text)
    response_3 = agent_3_iconographique(text)
    response_4 = agent_4_texte(text)

    # 3. Save results
    save_markdown_files(responses)

# Model: Ollama + qwen3:32b
# Timeout: 30 minutes per agent
        

Ollama

Inference locale

Local inference

Modele qwen3:32b sur mini PC local

qwen3:32b model running on local mini PC

Anti-hallucination

Regles integrees

Built-in rules

Validation des metadonnees depuis le contenu PDF uniquement

Metadata validation from PDF content only

Stack Technique

Tech Stack

Technologies utilisees pour la production

Technologies used for production

Technologies

Production PDF

PDF Production

LaTeX XeLaTeX Biber TikZ

Agents IA

AI Agents

Python Ollama PyPDF2

Presentation

HTML/CSS Playwright Reveal.js

Dev Tools

Claude Code Gemini CSS Variables

AI-HUB

Infrastructure locale NAS Local NAS infrastructure

Agents, modeles Ollama, configs Claude Agents, Ollama models, Claude configs

Resultat Final

Final Result

Le rapport final v2.6.0 - 34 pages (PDF)

The final report v2.6.0 - 34 pages (PDF)

Pages dans le rapport

Pages in the report

v2.6

Version actuelle

Current version

140+

Termes thesaurus

Thesaurus terms

Structure du Projet

Project Structure

LaTeX + HTML/CSS pour le PDF final

LaTeX + HTML/CSS for the final PDF

01-portfolio/latex/src/

main.tex                 # Point d'entree
├── content/
│   ├── 00-titlepage.tex    # Titre
│   ├── 01-carte-mentale.tex # Mind Map
│   ├── 02-demarche-recherche.tex
│   ├── 04-analyse-iconographique.tex
│   ├── 09-bibliography.tex
│   └── 10-annexes.tex
└── styles/
    ├── saintluc-config.tex # Config
    └── saintluc-commands.tex
          

main.tex                 # Entry point
├── content/
│   ├── 00-titlepage.tex    # Title
│   ├── 01-carte-mentale.tex # Mind Map
│   ├── 02-demarche-recherche.tex
│   ├── 04-analyse-iconographique.tex
│   ├── 09-bibliography.tex
│   └── 10-annexes.tex
└── styles/
    ├── saintluc-config.tex # Configuration
    └── saintluc-commands.tex
          

02-design-pages/

export-pdf.js            # Playwright
└── sources/
    ├── 01-couvertures/
    │   ├── page-garde.html
    │   └── quatrieme-couverture.html
    ├── 02-contenus/
    │   ├── analyse-iconographique.html
    │   └── methodologie.html
    └── 03-mindmap/
        └── MindMap_webproject/
            ├── mindmap-elixir.html
            └── export-png.js
          

LaTeX

Contenu textuel, bibliographie APA, structure academique

Text content, APA bibliography, academic structure

HTML/CSS/JS

Pages visuelles (couvertures, mindmap) exportees en PDF

Visual pages (covers, mindmap) exported to PDF

Pages du Portfolio

Portfolio Pages

Couverture & Sommaire

Cover & Table of Contents

Page 1 - Couverture

Page 1 - Cover

HTML CSS JS Playwright

Page 2 - Sommaire

Page 2 - Contents

LaTeX

Page 3 - Introduction

LaTeX

Pages du Portfolio

Portfolio Pages

Thesaurus & Methodologie

Thesaurus & Methodology

Page 7 - Mind Map

HTML CSS JS Mind Elixir Playwright

Page 8 - Methodologie

Page 8 - Methodology

LaTeX

Page 9 - Sources

LaTeX

Pages du Portfolio

Portfolio Pages

Analyse Iconographique

Iconographic Analysis

Page 15 - Grille d'analyse

Page 15 - Analysis Grid

LaTeX

Page 16 - Chagall

LaTeX

4eme de couverture

Back Cover

HTML CSS JS Playwright

Synthese

Summary

L'integration d'agents IA dans le workflow academique permet l'analyse automatisee de 40+ sources tout en maintenant la rigueur methodologique.

-- Denis Henin, Accenture Song

Integrating AI agents into the document workflow enables automated analysis of 40+ sources while maintaining methodological rigor.

-- Denis Henin, Accenture Song

Productivite Productivity
Analyse automatisee des sources Automated source analysis

Qualite Quality
Methodologie MDRD respectee MDRD methodology respected

Reproductibilite Reproducibility
Pipeline reutilisable Reusable pipeline

Autres Projets IA

Other AI Projects

Projets developpes avec Claude Code et agents IA

Projects built with Claude Code and AI agents

Projets presentes

Projects showcased

Multi-agents & Multi-modeles

Multi-agents & Multi-models

Autres Projets IA

Other AI Projects

Developpes en collaboration avec Claude Code et une multitude d'agents et modeles IA (1/3)

Built in collaboration with Claude Code and a multitude of agents and AI models (1/3)

News Intelligence

Agent de veille

News monitoring agent

Systeme multi-agents pour l'analyse automatisee de flux d'actualites.

Stack: Backend Python (FastAPI), frontend vanilla JS, base vectorielle pour recherche semantique.

Agents: Collecte RSS/API, analyse de sentiment, categorisation thematique, generation de resumes.

Output: Posts LinkedIn auto-generes, dashboard de veille, alertes personnalisees.

Multi-agent system for automated news feed analysis.

Stack: Python backend (FastAPI), vanilla JS frontend, vector database for semantic search.

Agents: RSS/API collection, sentiment analysis, thematic categorization, summary generation.

Output: Auto-generated LinkedIn posts, monitoring dashboard, personalized alerts.

Autres Projets IA

Other AI Projects

Developpes en collaboration avec Claude Code et une multitude d'agents et modeles IA (2/3)

Built in collaboration with Claude Code and a multitude of agents and AI models (2/3)

JurisAssist

Assistant juridique IA

AI Legal Assistant

Presentation de projet pour une application d'assistance juridique IA.

Contenu: Vision produit, architecture technique, roadmap de developpement, modele economique.

Design: Slides Reveal.js avec charte graphique professionnelle, animations fluides, responsive.

Livrable: Deck de presentation pour investisseurs et partenaires potentiels.

Project presentation for an AI legal assistance application.

Content: Product vision, technical architecture, development roadmap, business model.

Design: Reveal.js slides with professional branding, smooth animations, responsive layout.

Deliverable: Presentation deck for investors and potential partners.

Autres Projets IA

Other AI Projects

Developpes en collaboration avec Claude Code et une multitude d'agents et modeles IA (3/3)

Built in collaboration with Claude Code and a multitude of agents and AI models (3/3)

Equi SlideShow

Presentation client

Client presentation

Presentation client interactive pour un cabinet d'avocats specialise en droit du travail.

Format: Slideshow web responsive avec navigation tactile, mode plein ecran, export PDF.

Features: Animations CSS, transitions fluides, sitemap SEO, meta tags optimises.

Deploiement: Site statique heberge, temps de chargement <2s, score Lighthouse 95+.

Client presentation interactive for a law firm specialized in employment law.

Format: Responsive web slideshow with touch navigation, fullscreen mode, PDF export.

Features: CSS animations, smooth transitions, SEO sitemap, optimized meta tags.

Deployment: Static site hosted, load time <2s, Lighthouse score 95+.

Developpe en collaboration avec Claude Code et une multitude d'agents et modeles IA

Built in collaboration with Claude Code and a multitude of agents and AI models

Inspiration Design : Morag Myerscough

Design Inspiration: Morag Myerscough

Artiste britannique - Royal College of Art

British Artist - Royal College of Art

Morag Myerscough

Installations publiques depuis 1993

Public installations since 1993

Runway Gen-4

"Rends heureux ceux qui sont pres, et ceux qui sont loin viendront.""Make happy those who are near, and those who are far will come."

Son travail est a son meilleur dans les environnements brutalistes en beton. Des moments de pure joie en relief saisissant contre le gris impenetrable.

Her work is at its best in brutalist concrete surroundings. Moments of pure joy in stark relief against the impenetrable grey.

Formes geometriques bold

Bold geometric shapes

Triangles, cercles, rectangles

Triangles, circles, rectangles

Couleurs 100% saturees

100% saturated colors

Pas de pastels, energie pure

No pastels, pure energy

Art communautaire

Community art

Sentiment d'appartenance

Sense of belonging

Bonus : Creation du Logo avec Gemini

Bonus: Logo Creation with Gemini

Utilisation de l'IA generative pour le design

Using generative AI for design

Prompt envoye a Gemini

Prompt sent to Gemini

# AILE Logo - Gemini Prompt

Create a modern geometric logo for "AILE"
(AI-assisted LaTeX Engine)

DESIGN SYSTEM CONSTRAINTS:
- Colors: #FF6600, #FFE500, #0066FF,
          #FF0050, #00FF66
- Style: Bold geometric, no gradients
- Shapes: Circles, rectangles, triangles

CONCEPT:
- "AILE" = wing in French
- AI + Documents + Academic Research
- Reference: Martin Myerscough

OUTPUT: SVG 200x200px
          

Resultat Gemini

Gemini Result

SVG

Inspire par Morag Myerscough

Inspired by Morag Myerscough

Formes geometriques boldees, couleurs saturees, energie joyeuse

Bold geometric shapes, saturated colors, joyful energy

Animation "Dancing Shapes"

"Dancing Shapes" Animation

Chaque forme danse independamment - chaos joyeux organise

Each shape dances independently - organized joyful chaos

SVG anime natif

Native animated SVG

CSS keyframes, pas de librairie externe

CSS keyframes, no external library

Et vous ?

What about you?

Comment utilisez-vous l'IA dans vos projets ?

How are you using AI in your projects?

Denis Henin

Accenture Song

2025-2026

Questions Anticipees

Anticipated Questions

Annexe - Reponses aux interrogations frequentes

Appendix - Answers to frequently asked questions

Q&A

Processus & Methodologie

Process & Methodology

Combien de temps pour produire un rapport ?

How long does it take to produce a report?

~2-4h vs 2-3 jours manuellement. Le gain vient de l'automatisation de la recherche et de la mise en forme.

~2-4h vs 2-3 days manually. The gain comes from automating research and formatting.

Quelle est la part humaine vs IA ?

What's the human vs AI ratio?

~30% humain (direction, validation, ajustements) / 70% IA (recherche, redaction, formatage).

~30% human (direction, validation, adjustments) / 70% AI (research, writing, formatting).

Combien d'iterations avant un resultat satisfaisant ?

How many iterations before a satisfactory result?

Generalement 2-3 iterations. La cle : des prompts bien structures des le depart.

Usually 2-3 iterations. The key: well-structured prompts from the start.

Comment verifier la qualite du contenu genere ?

How do you verify generated content quality?

Validation croisee des sources, relecture critique, verification des citations et references.

Cross-validation of sources, critical review, verification of citations and references.

Questions Techniques

Technical Questions

Pourquoi Ollama en local plutot que l'API cloud ?

Why local Ollama instead of cloud API?

Confidentialite des donnees, cout reduit, independance, et latence plus faible pour les iterations rapides.

Data privacy, reduced cost, independence, and lower latency for rapid iterations.

Quelles specs machine pour faire tourner les modeles ?

What machine specs to run the models?

Mini PC: Ryzen 9 7940HS, 128GB RAM, 2x SSD 1TB. Heberge 24 modeles (459GB). 8B: 16GB min, 32GB confortable.

Mini PC: Ryzen 9 7940HS, 128GB RAM, 2x 1TB SSD. Hosts 24 models (459GB). 8B: 16GB min, 32GB comfortable.

Pourquoi LaTeX plutot que Word ou Google Docs ?

Why LaTeX instead of Word or Google Docs?

Controle typographique precis, reproductibilite, versioning Git, et separation contenu/presentation.

Precise typographic control, reproducibility, Git versioning, and content/presentation separation.

Quel modele LLM utilisez-vous ?

Which LLM model do you use?

Llama 3.2 (8B) pour les taches rapides, Mistral/Mixtral pour l'analyse complexe, Claude pour la redaction finale.

Llama 3.2 (8B) for quick tasks, Mistral/Mixtral for complex analysis, Claude for final writing.

Pratique & Ethique

Practical & Ethical

Est-ce acceptable dans un contexte academique ?

Is this acceptable in an academic context?

Oui, si transparent. L'IA est un outil, pas l'auteur. Toujours declarer son usage et valider le contenu.

Yes, if transparent. AI is a tool, not the author. Always declare usage and validate content.

Qu'en est-il des droits d'auteur ?

What about copyright?

Le contenu genere n'est pas protege. Verifier les sources, citer correctement, eviter le plagiat.

Generated content isn't protected. Verify sources, cite correctly, avoid plagiarism.

Quel est le cout de cette stack ?

What's the cost of this stack?

Ollama: gratuit. Claude API: $3-15/million tokens (selon modele). LaTeX: gratuit.

Ollama: free. Claude API: $3-15/million tokens (depending on model). LaTeX: free.

Comment apprendre a utiliser ces outils ?

How to learn to use these tools?

Documentation officielle, experimentations, communautes (Reddit, Discord), et surtout : pratiquer !

Official documentation, experiments, communities (Reddit, Discord), and most importantly: practice!

Limites & Perspectives

Limitations & Future

Quelles sont les principales limites ?

What are the main limitations?

Hallucinations possibles, dependance aux prompts, manque de creativite originale, besoin de validation humaine.

Possible hallucinations, prompt dependency, lack of original creativity, need for human validation.

Quand ne PAS utiliser cette approche ?

When NOT to use this approach?

Donnees sensibles/confidentielles, travaux necessitant 100% d'originalite, sujets tres recents (< 6 mois).

Sensitive/confidential data, work requiring 100% originality, very recent topics (< 6 months).

Quelles evolutions futures ?

What future developments?

Agents plus autonomes, meilleure integration multimodale, fine-tuning sur domaines specifiques.

More autonomous agents, better multimodal integration, fine-tuning on specific domains.

L'IA va-t-elle remplacer les humains ?

Will AI replace humans?

Non. Elle augmente nos capacites. La pensee critique, la creativite et le jugement restent humains.

No. It augments our capabilities. Critical thinking, creativity and judgment remain human.