TechnologyModels
FR
ProductivIA

Architecture & Models

Matania: sovereign orchestrator AI, agentic when needed

Matania is the intelligence that powers ProductivIA. It is itself an AI with vision, hosted in Canada, and directly handles the vast majority of requests, including those that require reflection and an agentic sequence: rewriting, summarizing, drafting, analysis, multi-step reasoning, visual reading and validation, with no external call. Its main practical limit is the available context window size. Beyond that, or when a highly specialized engine brings a real benefit, its multi-provider agentic capability comes into play: it calls and coordinates models from 12 AI providers and 16 specialized services, in an architecture independent of any single provider. Task decomposition, optimal model selection, reasoning loop, validation: everything is driven by Matania, transparently.

12 AI providers, one shared foundation

Activate, deactivate or prioritize each provider with a simple setting. Each model is monitored in real time (uptime, latency, costs, call volume).

hub

Conversation & reasoning

  • OpenAI: GPT-5.4 (plus mini, nano, pro), GPT-4o, ChatGPT Realtime
  • Anthropic: Claude Opus 4, Claude Sonnet 4
  • Google: Gemini 3 Pro, Gemini Flash, Gemma
  • xAI: Grok 4.1 (fast reasoning and non-reasoning)
  • Mistral: Large, Medium, Small, Codestral
  • Cohere: Command R+ and full library
  • Perplexity: Sonar (real-time online search)
palette

Media specialists

  • Google: Imagen 4 (image), Veo 3.1 (video), Lyria 3 (music)
  • xAI: Grok Imagine (image, image pro, video)
  • OpenAI: ChatGPT Image, GPT-4o-mini-tts, Sora (video)
  • Mistral: Voxtral (voice synthesis)
  • ElevenLabs: multilingual voices, dubbing, sound effects
  • getimg.ai: image generators (Flux and more)
  • Segmind: Kling, LTX, Seedance, Wan (video); Seedream, Ideogram (image); ACE Step (music)
shield

Sovereign (Matania)

Latest-generation open source models hosted on our sovereign servers in Canada, Gemma family augmented by our logic and knowledge layer:

  • Reasoning
  • Integrated vision: understands and evaluates images, screenshots, laid-out documents, generated applications
  • Multimodal (text, image, audio)
  • Lightweight models for routine tasks
  • Specialized models (code, translation, analysis)

This native vision makes the sovereign model the platform's internal judge: it can visually validate what other models produce, including external providers, and guarantee output quality. No data leaves our perimeter: you can restrict usage to sovereign models only if your policy requires it.

No vendor lock-in, by construction

The first layer of independence: Matania already covers most needs by itself, including in agentic mode, for tasks that require reflection or several steps. Other providers are only called beyond its scope, when a highly specialized engine brings a real gain, or when the context size requires it.

Beyond that, ProductivIA is not dependent on any AI vendor. If a provider changes prices, terms, limitations or usage policies, you can switch in one click to another one. If a new model appears that is more powerful or more economical, it is integrated and available immediately. This is strategic independence: not as a statement, but as architecture.

savings

Cost optimization

A simple rewrite does not need the same engine as a strategic analysis. Smart routing means costs can be reduced by up to 80% compared with exclusive use of a premium model.

balance

Ethical and geopolitical choice

Prefer to avoid certain providers for legal, ethical or geopolitical reasons? Disable them. Prefer to prioritize sovereign models? Configure it. Governance is in your hands.

How Matania chooses the model

Each request goes through several steps, all invisible to the user, all auditable by the administrator. Most requests stop at the first step: Matania answers by itself, without calling any external provider. The following steps trigger only for tasks that truly require a specialized engine.

analytics

1. Analysis and direct response

Matania evaluates the nature of the task: rewriting, analysis, reasoning, creative generation, code, multimodal. For everything it can process itself, including multi-step reflection within its context window, it answers directly, internally, without going further in the pipeline.

rule

2. Rule application

Institutional rules are checked: authorized providers, required confidentiality level, location constraints.

arrow_split

3. Model selection

The optimal model is chosen according to the cost / quality / time balance, while respecting the rules. If an external model is required, sensitive data is cleaned beforehand.

insights

4. Validation & iteration

Agentic loop: validation of the textual or visual result, with the sovereign vision model reviewing what other models produced, iteration if needed, and selection of a complementary model to verify or enrich. All under your control.

Sovereignty: three levels of depth

The institution chooses its level and can deepen it progressively.

cloud_done

Sovereign cloud

External models accessible according to your rules; sovereign models hosted on our infrastructure. Data is never hosted by a third-party hyperscaler.

dns

On-site server

All sovereign models run on your servers, on your premises. You can completely prohibit access to external models.

devices

Sovereign client

Workstations under native Boréal OS. Coupled with the sovereign server: a complete chain, from workstation to AI inference.

16 integrated specialized services

ProductivIA, through Matania, is not limited to text. The platform integrates 16 services that cover the full multimodal needs of a modern organization.

chat

Conversation & reasoning

Advanced chat, analysis, planning, autonomous agents, multi-step reasoning, contextual memory.

image

Image generation

Visuals, illustrations, diagrams, infographics, professional photos, multiple styles and formats.

record_voice_over

Voice synthesis

Natural multilingual voices, dubbing, narration, educational audio capsules.

hearing

Transcription & listening

Accurate audio transcription, speaker distinction, real-time analysis, used by the Meeting app.

movie

Video

Generation, editing, multilingual dubbing, element extraction, subtitling, video translation.

music_note

Music & audio

Music generation, sound effects, jingles, atmospheres for videos and presentations.

translate

Translation

FR/EN/ES/etc. plus First Nations languages. Respect for the original format, with tags and units preserved.

code

Code & data analysis

Code generation, statistical analysis, visualization, structured data processing.

search

Online search

Real-time web search, source verification, citations, thematic monitoring.

image_search

Vision & OCR

Image analysis, text extraction (OCR), scene description, reading scanned documents.

Agentic architecture: plan, execute, validate, iterate

ProductivIA is an agentic architecture, and Matania is itself the engine. For a complex task, it decomposes, plans, executes step by step, validates the result, and iterates if needed. As long as the task fits within its context window, Matania can run the entire loop by itself; it only mobilizes third-party services or models when context volume or sharp specialization justifies it. You stay in control and can intervene at any time.

route

Planning

Task decomposition into steps, identification of required services, execution order.

settings

Execution

Mobilization of the right services (text, image, video, code...), routing to the optimal models.

fact_check

Validation

Consistency checks, quality control, compliance with institutional rules, on text and visuals. Thanks to its sovereign vision model, Matania evaluates both its own deliverables (applications, documents, images, videos) and outputs from external providers.

refresh

Iteration

If validation fails or the user asks for an adjustment, Matania resumes, adjusts, recomposes.

An ecological approach by construction

Multi-provider is also an ecological choice.

Instead of systematically using an energy-hungry premium model for every task, Matania answers most requests itself, with no external call and no third-party tokens billed, and routes the others to the appropriate model. A simple rewrite consumes 100 times less than a complex analysis: why mobilize the same power?

energy_savings_leaf

Hydroelectric energy

Our infrastructure is powered by very low-carbon electricity in Canada.

network_intel_node

Smart routing

Light when that is enough, robust when necessary. No waste of cognitive resources.

cognition

Guided tools

Structured briefs consume fewer tokens than free-form conversations. Cognitive and energy savings.

Compliance & governance

verified_user

Law 25, GDPR, SOC 2 compliance

Compliance with applicable privacy protection laws, including Quebec's Law 25 and European GDPR principles, with SOC 2 certification in progress. A privacy impact assessment is performed before any out-of-province transfer.

admin_panel_settings

Logging & audit

Traceability for access and requests, consultable audit log, access controls by institutional silo.

delete_forever

Data control

Your data is never used to train external models. Erasable history, right to be forgotten, complete exports in open formats (see our philosophy).

block

Restriction to sovereign models

If your policy requires it, you can completely prohibit access to external models. All AI traffic stays on our sovereign servers, or on yours.

© ProductivIA 2026
info@productivia.ca - 581-504-0294
296, rue Saint-Pierre - Matane, QC G4W 2B9
Confidentiality Policy - Legal information