Labs

Before we offer a service, we build it in-house, on our own systems, and document what works and what does not. These eight labs are the basis of our recommendations.

Publication in progress. Detailed write-ups (architecture, method, results) are published as they are ready. We only publish a result once the measurement method has been described. Until then, we can walk you through the status of a lab in a conversation.

L1Being published

Measured sovereign RAG

An assistant that answers from internal documents, running entirely on controlled infrastructure.

L1 architecture: Measured sovereign RAG Documents indexed locally, a local model that answers, and an evaluation that measures answer quality. Documents Localindex Localmodel Answer quality measured with an evaluation set
What it demonstrates
That a document assistant can work without sending data outside, and that its quality can be measured rather than assumed.
What will be published
Architecture, evaluation method and measurement results.
Related services
Generative AI solution architecture · Private AI hosted on your premises
L2Being published

Enterprise AI gateway

A single entry point between staff and AI models: authentication, permissions, routing based on data sensitivity, audit log.

L2 architecture: Enterprise AI gateway Staff go through a gateway that authenticates, applies permissions, routes to the local or external model and records everything in a log. Staff Gatewayaccess, routinglog Local model External
What it demonstrates
That the use of several models can be governed in one place, with a record of who used what, when and with which data.
What will be published
Architecture, routing rules and a sample audit log.
Related services
Private AI hosted on your premises · AI governance and Law 25 compliance
L3Being published

Agent connected to office tools via MCP

An agent that acts in office tools (email, calendar, documents) through the Model Context Protocol, with limited permissions.

L3 architecture: Agent connected to office tools via MCP An agent acts in email, calendar and documents through the Model Context Protocol, with limited permissions. Agent MCPlimitedpermissions Email Calendar Documents
What it demonstrates
That an agent can be useful in existing tools without receiving more access than it needs.
What will be published
Architecture, permissions matrix and observed limitations.
Related services
Integrating AI with existing systems
L4Being published

The same RAG on Azure

The L1 solution rebuilt with Azure cloud services.

L4 architecture: The same RAG on Azure The L1 solution rebuilt with Azure services, then compared with the local version against a written grid. Documents Azure AISearch AzureOpenAI Answer compared with L1: quality, operations, data location
What it demonstrates
How to compare a local option and a cloud option honestly: quality, operations, operating costs and data location.
What will be published
Azure architecture and comparison grid with L1.
Related services
Generative AI solution architecture
L5Being published

LLMOps pipeline

Deployment, automated evaluation and monitoring of an AI application.

L5 architecture: LLMOps pipeline A prompt or model change goes through evaluation, deployment and monitoring, with rollback possible. Prompt ormodel Evaluation Deployment Monitoring
What it demonstrates
That a prompt or model change can be tested, deployed and rolled back like any other software change.
What will be published
Deployment pipeline, evaluation set and monitoring dashboard.
Related services
Production deployment and LLMOps
L6Being published

Penetration testing our own AI applications

Our own AI applications put through the attack categories of the OWASP Top 10 for LLM Applications.

L6 architecture: Penetration testing our own AI applications The attack categories of the OWASP Top 10 for LLM Applications are applied to our applications, which are then fixed. AttacksOWASP LLM AIapplication Findings Fixes
What it demonstrates
The testing method, typical weaknesses and how they are fixed, on systems we own.
What will be published
Method, anonymized findings and fixes applied.
Related services
AI system security
L7Being published

Law 25 register and PIA

The AI systems register and privacy impact assessments (PIAs) under Quebec's Law 25, written for NB-TECH's own systems.

L7 architecture: Law 25 register and PIA Each AI system is entered in the register, assessed through a PIA, then approved or adjusted by a written decision. AIsystem Register PIA Decision
What it demonstrates
That a small organization can keep its governance obligations up to date with simple templates.
What will be published
Register and PIA templates, with a completed example.
Related services
AI governance and Law 25 compliance · AI maturity assessment and roadmap
L8Being published

Industrial demand-response gateway

A gateway between industrial equipment and logic that curtails power consumption during peak periods.

L8 architecture: Industrial demand-response gateway A gateway reads industrial equipment and applies logic that curtails consumption during peak periods. Machines Gateway Curtailmentlogic Peakperiod
What it demonstrates
That AI can fit into an industrial environment by making recommendations, while fixed rules and people keep the final decision.
What will be published
Architecture, safety rules and decision log.
Related services
Intelligent process automation

Does a lab match your situation?

Every organization has its own systems, data and constraints. We always start by understanding your situation. Scope and investment are proposed afterwards, in writing.