We don’t ship every engagement starting from a blank repo. We ship from a
platform we built, and that we still build on, every day. It’s the
substrate that makes the governance, isolation, audit and self-hosting
story credible. And it’s licensable for teams who want to build on it
themselves.
Your cloud · Your data centre · Your controlSelf-hostedIsolated
What you buildThe platformData · isolated
Credit-Risk Workbench
Your application
KYC & AML Review
Your application
Whatever you build next
On the same foundation
Prototypical
/foundation - the application shell
Auth +Tenancy
DurableJobs
AuditBackbone
AI GatewayPer-call telemetry
Persona Factory
/personas - served over MCP
OpenWeave
/runtime - durable agent work
RLS enforced
PostgreSQL
Row-level security - isolation enforced here
Append-only audit trail
Export to your SIEM
Fig. 1One coherent foundation. Each product also stands on its own; Persona Factory serves anything that speaks MCP, and isolation is enforced at the database - not in application code.
/foundation
Prototypical
The foundation for AI applications that have to behave.
An application framework that deploys into your own cloud or data
centre, and supports multi-tenancy inside it, where the application
calls for it. The substrate every Katemba engagement is built on.
Database-enforced tenant isolation
PostgreSQL row-level security, applied table by table, with a
cross-tenant test suite that catches isolation breakage in CI
before it ever ships.
A registry architecture
Your application code stays cleanly separated from the framework -
upgrade the framework without rewriting your code, and your code
without rewriting the framework.
A durable job queue
The kind that survives restarts, retries the right things, and
tracks long-running work properly.
An event and audit backbone
Append-only, attributable, auditable. The substrate of every
governance story.
A provider-agnostic AI gateway
Anthropic, Azure OpenAI, Google Vertex, open-weights endpoints -
switchable per call, with cost and latency telemetry per provider.
An industry-standard stack
A modern web framework on PostgreSQL. Containerised.
Cloud-agnostic. Deploys into your environment.
Built for teams who need an auditable,
deployable-in-your-environment foundation with multi-tenancy ready
when the application demands it. The substrate that generic starter
kits skip.
/personas
Persona Factory
Version-controlled, evaluation-tested AI personas, served to anything that speaks the protocol.
A service that authors, versions, compiles, evaluates and serves AI
personas, skills, domain packs and output contracts.
persona/credit-reviewer
behaviour: assess credit memos against policydata_scope: tenant.credit_memos (read)+ boundary: never state a lending decision+ tool: policy_lookup (v2) tool: policy_lookup (v1)
EVAL · 142 CASES0.96PASS · PROMOTED
fig. 2: every version graded against its evaluation set. Promotion
is gated on passing. Illustrative values; the real runs happen in
your environment.
An authoring environment
Explicit behaviour, constraints, tools, data access, output
contracts: written down, not folklore in a prompt file.
Versioning and compilation
Every change is a version; every version is compiled into a
deployable artefact.
Evaluation runs
Every persona has an evaluation set; every version is graded;
promotion to production is gated on passing.
Governance metadata that travels with the persona
What’s allowed, what data it can see, who can change it.
Served over the Model Context Protocol (MCP)
Any compliant agent or assistant can consume governed personas
without one-off integration.
Built for organisations running more than
one AI assistant or agent who need them consistent, governed and
version-controlled instead of scattered across copied prompts.
/runtime
OpenWeave
Agents that survive a restart.
A runtime for long-running, resumable agent work, for workloads
where an AI task runs for minutes or hours and “start over from
scratch on failure” isn’t acceptable.
Durable runs
Agent work that survives process restarts, redeploys, and
infrastructure failures. The runtime remembers what was in flight
and picks it up.
Work claim queues
Multiple workers safely process the same pool of long-running
tasks without stepping on each other.
Checkpoints and plan graphs
Agents that work in discrete steps checkpoint progress and resume
from the last good point.
Memory
Agent memory primitives that are persistent, queryable and
bounded.
How you get the platform
Three routes in.
01
As part of a delivery engagement
Every Governed AI Delivery
and Vertical Workbench
engagement runs on the platform by default. You get the
application; the platform sits underneath, on your infrastructure.
02
Licensed for your team to build on
Platform Enablement
lets your engineers build governed, multi-tenant AI applications on
Prototypical, Persona Factory and OpenWeave themselves, with
Katemba architects alongside.
03
As individual components
For teams who want one piece - typically Persona Factory or
OpenWeave - alongside an existing application stack. Talk to us
about component licensing.
If you’re an engineering leader who wants to see the platform in motion
before talking commercially, we’ll set up a walkthrough with one of our
architects. No demo deck. A working environment and a code tour.
Want the code tour?
Tell us what you run today and we’ll walk you through the platform against it.