INDEPENDENT AI SYSTEMS ADVISORY · BERLIN

AI agents your organisation can govern.

I connect the business objective, agents, data, and delivery into a system your team can explain, operate, and evolve.

Discuss an initiative

From system design to delivery

/ 01

Three questions before AI enters real work.

01Which decision or workflow will this actually improve?
02Where may agents and data act — and where may they not?
03Who can explain quality, operations, and recovery with their own evidence?

/ SYSTEM

What we work on.

The hard part starts when specialised agents and teams must work together without receiving the same data, permissions, or responsibilities.

Decisions and workflows

Clarify the business objective, workflow, and measure before architecture and tool choice take on a life of their own.

Agent fleets and access control

Coordinate specialised agents across shared and restricted knowledge spaces, with distinct identities, role-based access, lane ownership, routing, and escalation.

Private ML and data systems

Design data paths, integrations, models, and human approvals so sensitive work stays inside a controlled boundary.

Delivery and operating evidence

Connect evaluation, release gates, observability, incidents, and recovery in a delivery chain the team owns.

One knowledge fabric. Different authorised views.

Relevant data is mapped together, classified, and stored with provenance. Each identity still receives only the view its role and task permit.

Mapped once: Documents, operational data, decisions, code, and evidence become one connected knowledge fabric instead of isolated context copies.
Restricted at retrieval: RBAC filters collections, documents, fields, and allowed actions before knowledge enters an agent context.
Actions authorised separately: Knowledge access grants no tool authority. Every call is checked again against identity, role, and expected effect.

GOVERNED KNOWLEDGE FABRIC / DEFAULT DENY

SOURCES MAPPED TOGETHERDOCUMENTSDATA + EVENTSTICKETS + DECISIONSCODE + EVIDENCE

MAP + CLASSIFYSource · owner · sensitivity

RBAC / POLICYIdentity × role × collection × action

PUBLICINTERNALRESTRICTEDPRIVILEGED

CENTRAL KNOWLEDGEStored together

READ + ROUTE

READ + WRITE FACTS

READ + LANE WRITE

EVIDENCE + VERDICT

COORDINATORPortfolio + objectives

DOMAIN AGENTDomain corpus

DELIVERY AGENTRepository + CI

REVIEW AGENTPatch + evidence

SEPARATE ACTION PATH

TOOL CALLIdentity + arguments

POLICY CHECKTool + expected effect

HUMAN APPROVALfor external effect

EXECUTIONOutcome logged

CENTRAL KNOWLEDGEStored together

PUBLICINTERNALRESTRICTEDPRIVILEGED

SOURCES MAPPED TOGETHERDOCUMENTS · DATA + EVENTS · TICKETS + DECISIONS · CODE + EVIDENCE

RBAC / POLICYIdentity × role × collection × action

COORDINATOR: View: Objectives, dependencies, and approved portfolio facts.Hidden: Domain documents, raw data, and protected repositories.
DOMAIN AGENT: View: Assigned domain collections down to document and field level.Denied: Data export, repository access, and other domains.
DELIVERY AGENT: View: Assigned repositories, CI evidence, and current revision.Denied: Other lanes, merge, release, and external effect.
REVIEW AGENT: View: Patch, tests, policies, and relevant evidence.Denied: Altering or approving the revision under review.

TOOL CALLIdentity + arguments
POLICY CHECKTool + expected effect
HUMAN APPROVALfor external effect
EXECUTIONOutcome logged

AUDIT LOGQuery · documents · decision · revision · tool outcome

The knowledge remains connected; access to it does not. Every query is filtered before context assembly. Tool authority is then checked separately, with human approval for external effect.

/ OPERATING MODEL

From initiative to a system you own.

Scope follows the initiative. One workflow needs a different depth from an agent platform spanning several teams.

The decision

Which operating constraint, risk, or decision makes the initiative worth pursuing?

The system

Which identities, knowledge spaces, permissions, interfaces, approvals, and dependencies form the real control boundary?

The operation

Which evaluation, release evidence, observability, and recovery does the team need before granting more authority?

Grounded in operating practice.

More than twelve years across engineering leadership, applied AI/ML, and secure on-prem and cloud data systems.

That includes operating whole agent fleets: multiple model providers and specialised roles, shared and restricted knowledge spaces, role-based access, lane ownership, structured hand-offs, exact-revision review, and recovery.

Damir MukimovFounder, Glowing Pixels — a software studio for AI and platform work, Berlin.

AI behind the firewallPrivate platform for documents, routing, and decision workflows Engineering platformPlatform work serving more than 200 engineers across 14 product teams

The work fits when product, security, and engineering hold different views of the same initiative, data or decision rights are unclear, or a pilot needs to become an operable system.

/ CONTACT