Private AI & Data Architecture
9 pieces from Duczer East's practitioners on running AI against data that cannot leave the premises: hybrid platforms, the compliance perimeter, and what a data layer has to deliver before an agent can act on it.
What this topic is about
Most enterprise AI is built on an assumption that holds for ordinary data and fails on the data a regulated institution cares about most: that the data will move to the model. Customer records, transaction history, clinical data and anything under a residency or contractual restriction do not go to a central cloud environment for inference, and the datasets that were treated as safely outside the AI perimeter because of that are now reachable by agents running where the data already lives.
The pieces in this collection work through what that reversal demands. Some are about the platform itself — hybrid architectures, the AI factory pattern, observability across on-prem and cloud, and where governance and lineage have to sit for the resulting system to be defensible. Others are about the layer underneath: why throughput and permissions are necessary but not sufficient, and why an agent acting on data that is stale, contradictory or means something different in the next system is a production failure waiting for its trigger.
Governing the agents that consume this data — identity, permission models, control planes — is collected separately in Agentic AI Governance & Model Risk.
Start here
The core argument: for the data that matters most in a regulated institution, the model has to come to the data, not the other way round.
-
Visibility Isn't Access: The Data Gap Where AI Stalls
Why the assumption that data will move to the model holds for ordinary workloads and breaks on the ones that decide.
7 min read -
Out of the Cloud Was Never Out of Compliance Scope
Agents can now work directly against on-premises data. The datasets treated as outside the AI perimeter no longer are.
3 min read -
Many Banks Already Run the Stack for Defendable Agentic AI
The document workflow and the data platform grew up in different parts of the bank. The seam between them is the architecture problem.
8 min read
What the data layer has to deliver
Throughput, permissions and freshness are necessary. None of them is sufficient if the data an agent acts on is wrong or means something different in the next system over.
-
AI Factories Solve Throughput. They Don't Solve Meaning.
Industrializing how fast data reaches models makes the unresolved question of what that data means the binding constraint.
4 min read -
Zero Permissions Won't Matter If the Data Underneath Is Wrong
Non-human identities at scale break the security model. Getting the permission model right is still only half the problem.
4 min read -
Agentic AI Is Not Just a Better Chatbot
What systems that act, rather than answer, require from data infrastructure: state, lineage, fallbacks, and integration built for machines.
6 min read
Notes from the field
Shorter reads on platform developments worth knowing about.
-
Cloudera and VAST Data Launch Unified AI Factory for Hybrid Enterprise AI
A production stack that answers the decade-old question of whether the AI comes to the data.
3 min read -
Native Observability Is the Heart of Hybrid Cloud
Unified telemetry across on-prem and cloud, and why the hybrid AI era makes it non-negotiable.
3 min read -
Cloudera and NVIDIA Got the Governance Architecture Right
Lineage and verifiable reasoning at the infrastructure layer rather than bolted on afterwards.
3 min read
AI on governed data, inside your perimeter
Duczer East builds private AI on Cloudera for regulated institutions — the platform, the data foundations and the lineage that let a use case clear model-risk and audit review.