How deeplinq is built
A four-layer architecture designed for sovereignty, auditability, and integration.
One engine. Every path runs through it.
Apps, agents and partners never reach your data or models directly. Every request passes through one headless engine. That's the single point where access, spend and audit policies are enforced, not just declared.
What "AI governance" actually means
It's one word that hides three pretty different things. Only the last one is something you can build into a product, and it's the part deeplinq owns.
01 · Societal & regulatory
The Macro Layer
The big "should AI even be allowed to do this" debate: the EU AI Act, risk tiers, safety evals, who's liable when something goes wrong. It matters a lot, but it's decided by regulators and society. It isn't something you ship in code.
02 · Organizational
The Policy Layer
Your own rules: who's allowed to use which model, on what data, for which purpose, and who's accountable for it. This part is real, but most of the time it lives in a slide deck or a policy doc that nobody can actually enforce.
03 · Technical & runtime
The Enforcement Layer
The mechanical part. Every single call the AI makes gets checked, allowed or blocked, attributed to someone, metered and logged, right when it happens. Not pieced back together in an audit three months later. This is the part deeplinq actually is.
The honest test: can the AI do something your policy forbids?
If it can, you've got a policy, not governance. If it can't, you've got governance. Because every request has to go through the engine, with deeplinq the answer is simply that it can't.
The four layers
Each layer is independently replaceable. The whole stack runs inside the perimeter the institution already controls, with RBAC and audit trail as cross-cutting concerns.
- Connector Hub
- Reads and writes to your existing systems : SAP, Oracle, Salesforce, Microsoft Dynamics, Temenos, Finastra, Murex, Avaloq, document stores, MES, SCADA, custom APIs. The platform meets your stack where it is — no data migration required.
- RAG Engine + Index
- Semantic indexing of documents, structured data, and conversational state. Hybrid retrieval combines keyword and semantic search, with source citations on every answer the agent returns.
- LLM Router + Orchestration
- Routes requests to the appropriate model. Cloud APIs (OpenAI, Anthropic, Mistral, Google) where the workload allows, or self-hosted open weights (Llama, Qwen, Mistral, Gemma, Falcon) where sovereignty requires it. Model versions are pinned ; multi-LLM is the default. Every call routes through one governed gateway — metered, audited, policy-enforced.
- Agent Orchestrator
- Where business teams interact via deployed agents — scoped per team, per use case, per role. Multiple agent types coexist : relationship-manager briefing, compliance research, operations triage, executive summary.
The same system, composed.
The layers above are the runtime view — how deeplinq governs. The build view is how you compose what ships: four primitives — Connectors, Agents, Apps, and Channels — built with one published SDK on the same public wire deeplinq's own teams use.
- Connectors
- Agents
- Apps
- Channels
Cross-cutting concerns
Three concerns wrap every layer.
Sovereignty
Four deployment modes : on-premise, air-gapped, your own cloud (AWS / Azure / GCP), or a deeplinq-managed multi-tenant cloud. The same platform binary runs in all four — sovereignty becomes a deployment choice, not a separate product.
Evidence
Append-only audit trail. Every prompt, retrieval, model call, routing decision, and agent action is hash-chained and signed. Tampering is detectable. Exports are formatted for the regulator, the DPO, and the internal auditor.
RBAC
Scoped access control by user, team, and agent. Document-level, record-level, and field-level controls. Retention, redaction, and residency travel with the access policy — one model, applied across the stack.
Deployment modes
The same platform binary, four envelopes. Choose by data class, regulatory regime, and CISO posture.
On-premise
Inside the institution's data centre. Models, vector stores, orchestration, and logs live on infrastructure the CISO already controls. For private banking, healthcare with strict residency, and public-sector workloads.
Air-gapped
No external network path. Inference, retrieval, and orchestration entirely on local infrastructure. Model updates and audit exports via controlled physical transfer. For classified workloads and offline industrial environments.
Your own cloud
Deploy into AWS, Azure, Google Cloud, or a sovereign-cloud region you choose. You own infrastructure, we provide platform — residency and data paths defined by your contract, not deeplinq's.
deeplinq-managed cloud
For workloads where the institution's compliance posture allows a managed-platform service. Tenant isolation, encryption, and evidence-trail discipline preserved at the platform layer.
Model agnosticism
The platform runs against the model that fits the institution's sovereignty and compliance posture. Cloud APIs for non-sensitive workloads — OpenAI, Anthropic, Mistral, Google. Open-weights models for on-premise or air-gapped deployments — Llama, Qwen, Mistral open, Gemma, Falcon. Switch per use case, per team, or per policy. Model-version pinning means the output produced today reconstructs the same reasoning twelve months later. The choice belongs to the institution, not the platform.
Integration with existing systems
Connector categories supported out of the box. Custom connectors built per engagement when the institution's stack requires it.
Core banking & wealth
Temenos, Finastra, Murex, Avaloq, Mambu, custom-built cores.
ERP & finance
SAP, Oracle, Microsoft Dynamics, Workday, NetSuite.
CRM & customer engagement
Salesforce, Microsoft Dynamics 365, HubSpot, custom CRMs.
Document & content
SharePoint, Box, document-management systems, internal archives.
Operational & industrial
MES, SCADA, AMOS, ARMS, FRMS, EFB, NMS, ticketing, change-control.
Messaging & channels
WhatsApp Business API, SMS (Twilio), email (Gmail, Microsoft 365 Graph), web, CLI — bound one-to-one to an app.
Search, vector & transcription
Milvus, Qdrant, Azure AI Search, OpenSearch, Typesense; two-stage rerank; transcription via Whisper.
Custom
Any system reachable by REST, GraphQL, ODBC, or message queue. Per-engagement scoping.
For full architecture review
We share the deeplinq architecture document, deployment-mode mapping, and integration catalogue under NDA with prospective institutional customers.