AI on your infrastructure
We deploy open models and tools on your servers or in a private cloud, connected to your documents and systems, with agreed acceptance tests and a handoff to your team.
A data boundary you can explain
Before deploying, we document where storage, inference, connectors, logs, backups and support access happen. We do not claim that no data leaves your network until the deployed system has been tested against that requirement.
Open choices
Open-source models and tools where they fit, evaluated for your case: licence, deployment behaviour and quality in French and English.
Integration with your tools
We connect AI to the systems you already use, rather than imposing a new platform.
A private AI assistant for confidential business documents
The most common case: finding information and getting cited answers across a large set of internal documents, without sending them to a public AI service.
A pilot starts with an authorized document set or sample data, never with sending us your documents at first contact.
- Search and cited answers across your procedures, contracts, reports and files
- Document summaries and comparisons, in French and English
- Separate access by team or by file
- A starting point for workflows: drafts, checklists, follow-ups
A document assistant helps find and summarize; it does not replace professional judgment. For high-stakes clinical, legal or financial decisions, human review remains necessary.
Sizing hardware for your workload
The right server depends on the model, the document volume, the number of concurrent users, text recognition (OCR) and the expected response time. We size it with your IT team or a partner.
- Workload estimate: models, concurrent users and document volume
- Choice between a workstation, a dedicated server or a private cloud
- Procurement and configuration coordinated with partners
Equipment warranty is the manufacturer's, under its terms. Software configuration, implementation and any support coordination by Neural Twiin are defined separately in the agreement. We do not publish hardware packages or prices.
The implementation process
Assess
An assessment of at least 30 minutes on the workflow, data and constraints.
Choose models and tools
Open models and tools evaluated for your task, your languages and your hardware.
Integrate and test
Connection to your systems and acceptance tests on an authorized set.
Deploy and hand off
Go-live, documentation and training. Model training only when justified.
What we check with you before go-live
- Answer relevance and citation accuracy
- Quality in French and English
- Access separation between users and teams
- Response time and concurrent users
- External-network behaviour of the deployed system
- Deletion, export and handoff to your operators
Responsibilities and exclusions
Neural Twiin
- Assessment, design and configuration
- Integration and acceptance tests
- Documentation and training
- Support under the agreed terms
Your organization or IT partner
- Choosing the authorized documents
- Operating the infrastructure
- Access and retention policies
- Assessments required by Law 25, where applicable
Excluded unless agreed in writing: hardware sales, warranty beyond the manufacturer's, on-site support, legal advice and model training (fine-tuning). Training a model is optional, only when justified.