Comuvia Advisory
Bring this reference architecture and its economics to your own build/buy, platform, or AI decision.
See advisory engagements →The system we advise you to build is the one we run
Every recommendation Comuvia Advisory makes is backed by a stack in production. This is the 2026 AI reference architecture we operate internally — and the same per-tier criteria, integration patterns, and cost economics we bring to your decisions.
A short overview of the Company AI System — the twelve-tier best-of-breed stack Comuvia runs, why compose it instead of a cloud bundle, and the honest trade-off.
Proof, not a slide
Most AI advice is theory. Comuvia's is the visible output of a running system.
Proof of execution
Comuvia operates a full on-prem AI and simulation environment. The people assessing your stack are shipping the same kinds of systems they assess.
Every artifact is its output
The media, research, and publishing across Comuvia's properties are produced by this system. The proof is something you can watch, not a reference you have to trust.
It backs every engagement
Architecture reviews, decision maps, fractional retainers, and simulation packs all draw on the same per-tier criteria and economics used to run it.
The reference stack
Comuvia composes best-of-breed systems tier by tier rather than buying a single cloud bundle. Each choice is made on its own merits and reviewed on a published cadence; the table below is descriptive of what we run, not an exhaustive configuration.
| Tier | What Comuvia runs | Why |
|---|---|---|
| Frontier LLM | OpenAI GPT‑5 family + Anthropic Claude, with additional open‑ and frontier‑weight models routed through an Azure AI Foundry gateway | Each call routed to the model that fits it — reasoning, editorial nuance, and cost optimized per task, not per vendor |
| Local LLM | Self‑hosted open‑weights models on Blackwell — fully local, no egress | Client‑confidential paths never leave the runtime |
| Agent framework | Anthropic Agent SDK + LangGraph | SDK for single-purpose agents; LangGraph for multi-step |
| Durable execution | Temporal | Mature; survival under crash and retry is critical |
| Image generation | On‑prem ComfyUI (SD3.5 / FLUX) for brand & cinematic control plus a cloud image model for subject‑grounded illustration | Two paths, routed by job: local control where it matters, high‑fidelity grounding at volume where it doesn't |
| Video generation | ComfyUI + Wan 2.2 + HunyuanVideo | Quality and speed on Blackwell |
| 3D / cinematic | Unreal Engine | Premium hero shots, MetaHuman, real-time |
| Physics / world-state | NVIDIA Omniverse + Isaac Sim (reserved — named, not yet in production) | Simulation, validation, synthetic data |
| Decision simulation | Dyno-Sim (59 scenario types, stock-flow-consistent families) | Accounting-identity discipline; behavioral, monetary, demographic depth |
| Content origin | BookWriter | Source-tracked corpus, chapter drafting, reference discipline |
| Publishing ops | MediaManager | Platform adapters across owned sites and channels |
| Shared governance | Shared Services | Tier-based approval gates, eval pass before publish, cost ledger + audit trail |
Why best-of-breed, not a cloud monolith
The major cloud vendors all market a one-stop substitute for the table above. Comuvia chose composition deliberately, for four reasons.
Cost optimization across tiers
Frontier-LLM economics and visual-generation economics move on different curves. Composing the stack lets each tier be optimized on its own cost basis instead of one bundle's pricing.
No lock-in for client deliverables
Recommendations and artifacts stay portable. Nothing about an engagement forces a client onto a provider Comuvia happens to resell — because Comuvia resells nothing.
Sovereign data paths
A self-hosted local model keeps client-confidential work inside the runtime. Sensitive paths never have to traverse a third-party API.
Multi-model routing
GPT‑5 for reasoning-heavy drafting, Claude for editorial nuance, a local open-weights model for confidential work, fast media models for throughput — each task routed to the system that fits it best, and re-routed as the frontier moves.
The honest trade-off: operational complexity is higher than a monolith. Best-of-breed composition makes sense when running this stack is your business; it does not always make sense for an enterprise whose business is something else. Which path fits your situation is exactly the kind of question an Architecture & Decision Review Sprint is built to settle.
Where to go next
Governance and provenance
The approval tiers, eval discipline, and provenance posture that govern everything this system produces.
Read the governance posture →