Our products
in action.
Your challenges.
Our solutions.
Problem
Teams want to use ChatGPT, Claude or Gemini for daily productivity. But each provider has limitations, costs, terms. No visibility on usage. Impossible to switch models without starting over. Shadow AI spreading.
Our R&D commitment is recognized by French innovation authorities.
BFTe Grant Recipient
Bpifrance
French Tech Emergence Grant recipient and DeepTech labeled. Breakthrough technology from research.
JEI Status
Ministry of Higher Education
Official recognition for companies dedicating over 25% of expenses to R&D.
Systematic Cluster
Competitiveness cluster
European DeepTech hub: AI, cybersecurity, digital infrastructure, photonics.
France DeepTech
National collective
Network of 300 startups, funds and labs committed to technological sovereignty.
The challenge.
In numbers.
of generative AI projects never reach production.
Source: MIT, 2025
of companies abandoned most of their AI initiatives in 2025. Compared to 17% in 2024.
Source: S&P Global
Why does it fail?
Data
Poor quality, heterogeneous formats, no structured pipeline.
Technical
No MLOps expertise, too complex stack, external dependencies.
Costs
Working prototype, but impossible to scale without blowing the budget.
Value
No clear use case, ROI difficult to demonstrate.
The challenge is never the proof of concept.
It's making it to production.
Sovereign across the stack.
From data to software.
From dataset to deployment: collection, annotation, training, inference, integration. One team, end to end.
Full control
From infrastructure to models, we develop and operate every component in the chain.
Unified architecture
A coherent stack, no third-party dependencies. Less complexity, more reliability.
Industry focus
Energy, defense, industry: solutions designed for your regulatory and operational requirements.
Our technical
approach.
Hybrid Architecture
Open-source and proprietary models depending on the use case. Cloud, on-premise GPU, or both. We choose the best solution for each need.
Controlled Infrastructure
PostgreSQL, vLLM, pgvector. You keep control of your data and your stack. No vendor lock-in.
Flexible Deployment
SaaS, on-premise, air-gap. We adapt to your security and compliance requirements.
Measurable Results
Public benchmarks on HuggingFace, performance metrics, evaluations on your data. Full transparency.
Tech Stack
From data
to model.
Our complete pipeline: from your raw documents to a specialized model deployed in production.
Dataset Creation
Internal Documents
PDF, images, technical drawings
OCR / Extraction
VLM for structured extraction
Annotation
Human or synthetic Visual QA
Business Logic
Domain rules injection
Dataset
Multimodal JSONL format
Training
Base Model
Open-source or proprietary VLM
Configuration
QLoRA / LoRA adapters
Training
Supervised fine-tuning
Fusion
Adapter weights merge
Inference
Specialized model deployed
Our capabilities
in action.
From document to drone, same perception engine. Our Vision-Language Models (VLM) are at the heart of all our capabilities.
Legacy applications without API? Autonomous post-RPA navigation.
SAP, Oracle, in-house ERP. Our agents navigate your interfaces like a user, without API integration.
Features
- UI-DETR: interface element detection
- Screenshot-by-screenshot navigation
- Beyond traditional RPA
Agent in action — Autonomous navigation.
Computer Use / Post-RPA
Legacy applications without API? Autonomous post-RPA navigation.
SAP, Oracle, in-house ERP. Our agents navigate your interfaces like a user, without API integration.
- UI-DETR: interface element detection
- Screenshot-by-screenshot navigation
- Beyond traditional RPA
Field / Inspection
Camera, drone, video stream. Real-time autonomous decisions.
Anomaly detection, quality control, infrastructure inspection. Autonomous navigation, obstacle detection, embedded decision loops.
- Real-time video stream
- Autonomous detection and navigation
- Models trained on field data
Fine-tuning / Specialization
Models trained on your data.
Supervised fine-tuning, custom datasets, business evaluation.
- Training on your data
- Annotated datasets
- Continuous evaluation
Documents / Extraction & RAG
PDFs, technical drawings, diagrams. Smart extraction and search.
Our VLMs analyze and index your documents. Precise extraction with exact source, semantic search in natural language, sourced answers.
- Domain-specialized VLMs
- Structured JSON extraction + vector indexing
- Multi-format: PDF, Word, Excel, CAD
- Answers with citations and sources
Agents / Autonomous workflows
Intelligent pipelines that adapt in real-time.
Our agents orchestrate complex end-to-end workflows: ingestion, extraction, validation, enrichment. Each step is autonomous, with contextual decision-making and automatic correction loops.
- Multi-step pipelines with conditional branching
- Automatic validation and enrichment
- Real-time monitoring and detailed logs
- ERP, CRM, database integration
Robotics / Autonomous systems
Drones, robots, autonomous machines. Real-time embedded VLM.
Perception, autonomous navigation, embedded decision-making. Our models run directly on edge hardware with minimal latency.
- Compact VLMs for edge inference (500M-2B)
- Optimized real-time latency
- NVIDIA Jetson integration, industrial hardware
Four products.
One goal.
Pi-Search
Private CloudTurnkey application for your teams. Document search, extraction, generation. Simple interface, VLM power.
- Cloud-agnostic hosting
- Deployments up to thousands of users
- All your documents indexed
- Model specialization and training on your data
- LLM-agnostic (provider choice)
Pi-Edge
On-PremiseComplete deployment on your infrastructure. Models, inference, data. Everything stays with you. Air-gap compatible.
- Installation on your servers
- On-premise or private cloud GPU
- No outgoing data
- Team training included
- Priority support
Pi-Computer
Post-RPAAutomation of applications without API. Our agents navigate your legacy interfaces like a human user.
- Autonomous SAP, Oracle, ERP navigation
- Intelligent UI detection (UI-DETR)
- Exception and variation handling
- Alternative to traditional RPA
- On-premise deployment
Pi-Robot
Edge / EmbeddedEmbedded VLM on your autonomous machines. Drones, robots, industrial equipment. Local decisions, real-time.
- Edge-optimized models
- Offline operation
- Real-time latency
- Hardware-agnostic
- Integration SDK
Need help choosing? Let's discuss your context
Open research.
Published models.
Our work is public. Models and datasets available on HuggingFace.
Models & Datasets
Publications
European VLM for Nuclear
Fine-tuning SmolVLM for nuclear regulatory documents. 15x improvement in French.
European Sovereignty with Flantier
European alternative to Chinese models. 76.7% accuracy vs 19% baseline.
UI-DETR-1 for Autonomous Agents
Open-source UI detection (MIT). +20% vs OmniParser. Unrestricted commercial use.
Let's discuss
Your Project.
Document AI, legacy automation, field inspection. We deploy solutions that make it to production.