Architecture & attention engineering
Agent runtimes, governed inference, and tool mediation. Deciding what an agent sees now and what it retrieves later. Implemented through Clawdapus and cllama.
AI architecture · Technology leadership
Wojtek Grabski
I design how AI systems decide what to attend to, what to remember, and what they are allowed to do. I lead the teams that put those systems into use.
At Flux Inc., I connect AI architecture and hands-on engineering with product direction, operations, and company-wide adoption. My broader experience includes leading teams of up to 50 people and taking products from concept through deployment and commercialization.
- Based in
- Marietta, Georgia
- Company
- Founder, Flux Inc. · 2007–present
- Contact
- [email protected]
AI architecture & practice
Context management, memory, retention & recall
Role-aware context, explicit priorities, source-linked retention, and selective recall. Company context in My AI; recall and retention hooks in cllama.
Governance & safety
Scoped permissions, human review, and application-enforced actions. Model-governance design; supervised support drafting; human-gated, time-limited access to operational systems.
Operations
Deployment, fleet management, and AI toolchain distribution across customer instances. Shared-work coordination through Talking Stick; internal company tooling for task ownership, review, handoffs, and recovery.
Evangelism & adoption
Staff demonstrations, training, onboarding, and shared tools. Connecting AI capabilities to daily work, and helping people direct, review, and correct the systems.
AI & context
Principles I use when designing these systems.
I treat context assembly, memory, tool execution, and authorization as separate system responsibilities. The model interprets the task; the surrounding software controls what it sees, what it can execute, and what survives the session.
01 / Now
Present
Supply role, goal, current state, authority, and constraints before the agent acts.
02 / As needed
Look up
Retrieve documents, records, and current system details for the task.
03 / After work
Retain
Record decisions, preferences, outcomes, and their sources for later use.
04 / Next task
Recall
Select retained information relevant to this person, role, and purpose.
Selection and access depend on role, purpose, and authority.
Context & attention
Assemble context for each invocation
I select context by role, goal, work state, and authority. Essential state is supplied before inference; detailed records remain available through tools. Attention engineering means controlling relevance, priority, placement, and volume within that context.
In cllama, the proxy assembles feeds, context blocks, and recalled memory. Context snapshots and prompt hashes make the assembled input inspectable. RailsTrail exposes possible next actions; application checks still decide whether an action can execute.
Memory
Separate source records, retention, and recall
I keep source history distinct from derived decisions, preferences, questions, and outcomes. Derived records need provenance and explicit priority so they can be corrected or reinterpreted. Recall selects for the current task; it does not replay the entire archive.
cllama provides pre-turn recall and post-turn retention hooks, bounded memory blocks, and source/kind filtering. A failed recall adds no memory; it is not treated as a successful lookup.
Culture
Represent priorities and their scope
I treat organizational context as more than document retrieval. It includes which outcomes matter, who makes decisions, what can be disclosed, and how information is explained to a particular audience.
People express this knowledge in ordinary language. I preserve the source instruction and distinguish personal, team, and organization scope. Changes that affect other people require the appropriate authority; unresolved conflicts go back to someone authorized to decide.
Governance & safety
Enforce authority outside the model
Instructions inform the model; they do not grant access. Authentication, data access, and action authorization belong in the application or execution layer. Preferences guide responses and can change with context. Permission changes require an authorized decision.
My AI checks proposed changes against policy from a trusted base revision. Internal operational access is human-gated and time-limited, with expiry, revocation, and an audit trail. Model-based policy evaluation supplements these controls; it does not replace them.
Runtime & tools
Declare contracts and bound execution
I treat agents as untrusted workloads. Model selection, identity, tools, skills, feeds, and memory are explicit runtime inputs. Tool execution needs limits and defined failure behavior, independent of how confidently a model requests an action.
Clawdapus compiles agent declarations into Docker/Compose artifacts and runtime manifests. cllama bounds managed tool loops with round limits, per-tool and total deadlines, duplicate-call replay, and repeated-call cutoffs.
Work orchestration
Persist work state independently of sessions
Task ownership, review, blockers, and handoffs belong in durable work records. A replacement session reconstructs the current state from those records. Completion needs evidence of the result, not just a successful process exit.
Talking Stick coordinates shared writers through leases and structured handoffs. Internal company orchestration tooling tracks tasks, assignments, review, and recovery across human and agent work.
Culture, authority, and memory
Internal design discussion · April 30, 2025
An edited excerpt from my discussion of organizational context, scoped authority, and selective recall, before work began on the model-governance layer. It records ideas being explored at the time.
Edited excerpt · 6 min 37 sec
Projects
Tools related to these principles.
Original design with AI-assisted implementation. Public projects and internal company tooling.
RailsTrail ↗
API descriptions of possible next actions, with execution checks in the application.
PublicWork orchestration
Internal company tooling for durable tasks, assignments, human review, handoffs, and recovery across agent sessions.
InternalOperational access
Human-gated, time-limited access to operational systems, with expiry, revocation, and an audit trail.
InternalCareer history
Overlapping work, not consecutive jobs.
2025–present
AI implementation
Flux Inc.
I designed a model-governance layer around roles, memory, authority, and company rules. The team developed retrieval, conversational rules, and compliance mechanisms.
My implementation work includes a DICOM MCP prototype and AI-assisted reconciliation of purchase and installation records. In 2026: Clawdapus, cllama, My AI (my-cli), staff training and onboarding, and supervised support drafting. My work also includes fleet management and distributing the internal AI toolchain across customer fleet instances. Internal company tooling covers durable work orchestration and human-gated, time-limited operational access.
2007–present
Founder
Flux Inc.
Medical-imaging software and integration. DICOM Printer was sold and integrated into imaging workflows in 2007. Our products have served thousands of facilities through instrument vendors and hospital networks.
I created DICOM Capacitor and contributed to DICOM Printer 2. I lead product development, customer delivery, and commercialization. Technical work includes distributed image delivery, reconciliation, redundancy, recovery, and customer support.
2013–2017
Product design & delivery
Product and client work
Radiology reporting software, content and commerce applications, and pharmacy integration design. DXA Desk rollout planning covered training, test exams, paper fallback, and clinical acceptance. At Informedy, I owned the product from proposal through implementation and handoff.
By 2010
Co-founder
Influx
I built the clinical product from intake and scheduling through reporting, billing, and financial reconciliation, including the staff workflows around it. I developed Influx’s entire internal software stack. I also designed the branding and marketing, managed trade shows and sales channels, and did board-level work.
2009 & 2011Documented work
Production systems
Yellow Pages, through Flux
I managed a multidisciplinary team of developers, voice artists, and copywriters, and built the production system around the work. It covered high-volume website and media production: intake, copy, design, staffing, quality checks, and delivery. Responsibilities included change control, capacity planning, quality assurance, and staff training.
2000s
Research computing
Affect and Cognition Lab
University of Toronto
Research computing, imaging, and experimental systems in cognitive neuroscience. I worked on the mathematics of early facial-expression machine-learning experiments with students from Geoffrey Hinton’s lab.
1995–2004
Software & IT
Aymes Medical
Medical reporting software, installation, IT management, and technical, customer, and sales support.
1995–1996
Business databases
Hess Packaging
Database development and production management systems.
Background
Engineering & business systems
C#, C++, Ruby, Python, Go, Rust, JavaScript, TypeScript, SQL, and shell; DICOM and HL7 interoperability. Systems built include staffing and forecasting applications, enterprise management, production management, and clinical delivery.
Studies
University of Toronto: applied physics, focused on photonics and condensed matter, with signal processing and digital imaging. Earlier studies in mineral engineering. No degree awarded.
Publications
Expressing fear enhances sensory acquisitionNature Neuroscience · 2008
Emotional memories are not all created equal: evidence for selective memory enhancementLearning & Memory · 2006