10-week engineering internship at Cavalons, a French equestrian tech startup, focused on designing and integrating an AI-powered coaching chatbot for the ÉquiProgress application.
| Field | Value |
|---|---|
| Student | Manil Doudou |
| School | CESI — Y3 Computer Science Engineering |
| Company | Cavalons (Paris, Île-de-France, France) |
| Period | January 12, 2026 – March 8, 2026 |
| Company Tutor | Ms. Camille Despesse (CEO) |
| Academic Tutor | Mr. Wassim Masri |
Cavalons is a startup that connects horse owners and riders through a peer-to-peer platform (demi-pension, horse sitting, experience sharing). During my internship, the company was building ÉquiProgress — an AI coaching assistant for equestrian training.
My primary mission was to design, build, and integrate the core AI chatbot for ÉquiProgress, covering:
- Conversational flow design on Botpress
- Integration with the Google Gemini 1.5 Pro API
- Equestrian knowledge base curation (RAG architecture)
- Data analysis of equestrian datasets (Cavalons, Ekwo, Appaloo)
- Supabase backend schema design (13 interconnected tables)
- Embedding the chatbot into the Figma mock-up prototype
- Strategic research on AI Agents and AI-powered training generators
equiprogress-ai/
│
├── docs/
│ ├── reports/
│ │ ├── Discovery_Report_Manil_Doudou.pdf # Mid-internship report (week 4)
│ │ └── Final_Report_Internship_Manil_Doudou.pdf # Full 10-week final report
│ └── analyses/
│ ├── Comparatif_Appaloo_EKWO.pdf # Comparative dataset analysis
│ ├── Analyse_des_données_Cavalons.pdf # Cavalons dataset deep-dive
│ ├── Analyse_des_données_Ekwo.pdf # Ekwo dataset deep-dive
│ └── Liste_des_base_de_données.pdf # Database inventory
│
├── data-analysis/
│ ├── notebooks/
│ │ ├── cavalons_data_analysis.ipynb # EDA on Cavalons horse dataset
│ │ └── analyse_ekwo_db.ipynb # EDA on Ekwo training database
│ └── datasets/
│ ├── chevaux_equiprogress.xlsx # Horse profiles dataset
│ ├── Excel_donnees_ekwo_final_uniforme.xlsx # Standardized Ekwo data
│ └── Excel_Chatbot_Cavalons.xlsx # Chatbot knowledge base (Excel)
│
├── ai-chatbot/
│ └── Plan_action_Chatbot_Cavalons.pdf # Full chatbot development plan
│
├── research/
│ └── Mission1_Agents_IA_Powerpoint.pptx # Study of AI Agents (Mission 1)
│ └── Mission2_Agents_IA_Powerpoint.pptx # Study of AI Agents (Mission 2)
│
└── README.md
| Layer | Technology |
|---|---|
| Conversational AI | Botpress — NLU, dialogue flows, KB |
| Language Model | Google Gemini 1.5 Pro API |
| Backend / Database | Supabase (PostgreSQL, Auth, Edge Functions) |
| Frontend | React Native + Figma mock-up |
| Data Analysis | Python — pandas, matplotlib |
| Project Management | Trello, Google Meet, WhatsApp |
The first phase of the internship involved exploratory analysis of three equestrian datasets to extract training patterns, identify recovery protocols after veterinary interventions, and benchmark existing market data.
| Dataset | Source | Records | Description |
|---|---|---|---|
| Cavalons | Internal | ~1,000+ | Horse profiles, age/height distributions |
| Appaloo Exercices | External | 470 | Equestrian exercises (15 fields) |
| Ekwo Entraînements | External | 77 | Training sessions (6 fields) |
- Appaloo offers 6× more exercises with richer metadata (15 fields vs 6), strong in Obstacle (131) and Cavalettis (95)
- Ekwo is more beginner-friendly (55% beginner/easy vs 28% for Appaloo), stronger in Dressage (31) and CSO (16)
- The two databases are complementary and were recommended for harmonization
- Recovery pattern analysis after veterinary interventions (injections, osteopathy) produced evidence-based convalescence guidelines embedded into the chatbot's logic
→ See data-analysis/notebooks/ for the full Python analysis.
The ÉquiProgress AI coach was built with a layered architecture designed for safety, personalization, and pedagogical soundness.
User Input
│
▼
[Botpress NLU] ──── intent classification + entity extraction
│
▼
[Context Gathering Flow] ──── horse name, energy, duration, objective
│
▼
[Supabase API Call] ──── user profile, last 5 sessions, health events
│
▼
[Gemini 1.5 Pro API] ──── system prompt + context + knowledge base (RAG)
│
▼
[Safety Filter] ──── risk-tiered response guardrails
│
▼
Structured Response: warm-up → main exercise → alternatives → calendar prompt
The chatbot implements a risk-tiered response system:
- Low-risk (fun flatwork exercise): high AI autonomy, creative suggestions
- Medium-risk (horse seems stiff): conservative suggestions, vet reminder
- High-risk (horse is sick/lame): guardrail flow, no strenuous exercise, mandatory professional consultation
The chatbot is strictly scoped to equestrian topics and will never attempt to diagnose injuries.
Two research presentations were delivered to Cavalons leadership:
- Typologies of AI agents (reactive, deliberative, hybrid, LLM-based)
- Modern agent architectures: memory, RAG, tool use, planning
- Comparative analysis of tools: LangChain, LlamaIndex, CrewAI, Botpress, N8N, Vertex AI
- Comparative analysis of LLMs: GPT-4o, Gemini 1.5 Pro, Claude 3.5, Mistral
- Case studies: Petriage, Happieanimals
- Model selection for text + image generation
- Exercise source strategy (legal, quality-controlled)
- Personalisation via RAG + vector databases (pgvector)
- Proposed 3-phase technical architecture: rule-based → hybrid RAG → multimodal
→ See research/presentations/
| Document | Description |
|---|---|
| Discovery Report | Analysis of Cavalons' structure, culture, and early technical contributions (Week 4) |
| Final Internship Report | Comprehensive 10-week report covering all 4 phases of the internship |
# Clone the repository
git clone https://github.com/YOUR_USERNAME/equiprogress-internship.git
cd equiprogress-internship
# Install Python dependencies
pip install -r requirements.txt
# Launch Jupyter
jupyter notebook data-analysis/notebooks/- AI Safety is non-negotiable — building a domain-specific guardrail system was as important as the chatbot itself
- RAG over fine-tuning — for a small startup, grounding the LLM in a curated knowledge base is faster, cheaper, and safer
- Integration is the hard part — connecting Botpress → Supabase → Gemini required deep knowledge of three separate systems
- Domain knowledge matters — understanding equestrian training principles was essential to build a credible product
Copyright © 2026 Cavalons. All rights reserved.
This software is proprietary and confidential. Unauthorized copying, distribution, modification, or use of this software, in whole or in part, is strictly prohibited without prior written consent from Cavalons.