Werner Kasselman
Enterprise Architect · Founder, Verivus
I build the evidence and governance layer for AI agent systems: semantic code graphs, machine-checkable plans and multi-model review gates that decide what ships.
Gold Coast, Australia.
Current work
My work now is AI agent infrastructure. I design the systems, direct the agent swarms that write the code, and own the review gates and evidence standards that decide what ships; the open source and upstream work below is the public part of that.
At Verivus, I'm working on a governed platform for business decisions and actions, customer-controlled access from AI agents to tools and data, and developer tools for understanding code and coordinating coding agents. The Assurance Platform and hosted AgentFederator service are in development; sqry, llm-cli-gateway and agent-assurance are released as open source.
Open source I built
The three main pieces cover understanding code, coordinating coding agents, and checking the plans and evidence those agents work from.
-
sqry: semantic code search
A Rust code search engine that parses code into a graph over tree-sitter ASTs, so an agent can ask who calls what across 37 language plugins; it runs as a shared daemon with MCP, LSP and VS Code integrations, ships MIT-licensed releases with artifact attestations, and installs from a Homebrew tap. It surfaced the four kernel fixes listed below.
-
llm-cli-gateway: one control plane for coding agents
A local MCP control plane that runs Claude Code, Codex, Gemini, Grok, Mistral Vibe, Devin and Cursor agents behind one interface, with durable async jobs, session continuity, workspace-scoped execution, approval gates and signed MIT-licensed releases; it carries the cross-model review gate that substantive changes pass before they merge.
-
agent-assurance: machine-checkable planning for agents
An Apache-2.0 specification for writing agent plans as dependency DAGs in TOML, with traceability and review-readiness proofs that a validator recomputes; the companion repository holds the manuscripts and executable proof bundles.
More open-source work (2)
-
llm-tips: a writing standard for AI-assisted prose
An MIT-licensed standard that explains why instruction-tuned models write the way they do, from the grammatical shifts Reinhart, Brown et al. measured (PNAS 2025) to the em-dash rates Freeburg measured across twelve models, and turns that into density rules scored across a whole piece; it ships a 56-rule machine-readable policy, a linter that lists the rules it could not check, and a proposal-triage checker, with 43 passing tests.
-
agent-contribution-assurance-pack: CI governance for AI-assisted pull requests
An MIT-licensed pack that open-source projects install to gate AI-assisted pull requests with bash and GitHub Actions alone: it sorts each change into three tiers from the files it touches, fails a change unless its research summary cites a file the pull request actually changes, and requires high-impact changes to commit a tier-matched plan in the same pull request; 79 validator cases pin that behaviour.
Upstream contributions
-
Four fixes merged to the mainline Linux kernel
Two ksmbd memory-safety fixes, for a use-after-free and NULL dereference in smb_grant_oplock() and memory leaks and a NULL dereference in smb2_lock() (Linux 7.0, marked for stable), and two DRM integer-overflow fixes in radeon_align_pitch() and amdgpu_gem_align_pitch() that stop unprivileged userspace allocating a zero-size dumb buffer (Linux 7.2, radeon also queued for 7.1 stable). sqry call-graph analysis surfaced all four.
-
Chromium: scratch-buffer copy for Slug deserialization (CVE-2026-113078)
The fix copies Slug input from renderer-controlled shared memory into scratch space with CopyScratchSpace(data_bytes) before Slug::Deserialize() and removes the const_cast that stripped volatile from the shared-memory pointer.
-
NVIDIA SkillSpector: security scanner for AI agent skills
Three merged pull requests to NVIDIA's open-source scanner: #6 made the MCP tool-poisoning parameter checks reachable on real scans, #133 made LLM finding schemas portable to structured-output endpoints, and #375 made the CLI report the findings that drove the risk score. A commit on main credits me with finding an SSRF allowlist bypass through redirects, and issues #9 and #11 I reported on Stage 2 LLM resilience were fixed upstream.
More upstream contributions (5)
-
Celesto (formerly SmolVM): nine merged pull requests
Ten pull requests over two days in August 2026 to an Apache-2.0 sandbox platform for AI agents with Firecracker, QEMU and libkrun backends; nine merged and I withdrew the tenth. The merged work fixed an image cache check fooled by coarse directory timestamps (1,924 of 2,000 trials), a liveness probe that misread pid 0 and negative pids, a blocking hypervisor call on the async event loop and network metadata that did not match the guest, each with deterministic tests.
-
Crossplane: reconciler panic root-caused (#7479)
Root-caused a nil-dereference panic in the Crossplane package runtime reconciler that fires when an ImageConfig's optional pull-secret reference has gone away; a Crossplane maintainer acknowledged the report and asked for the fix as a pull request.
-
gstack: security audit fixes landed with co-author credit
Two audit rounds of Garry Tan's gstack, run with sqry and llm-cli-gateway, produced ten issues that were all closed as completed, and Garry took the new fixes from my two audit pull requests into main in one security-wave commit (PR #847, April 2026) with co-author credit under mr-k-man, this account's former GitHub login; the pull requests themselves were closed once the fixes landed.
-
chrome-agent (fork): Windows support
On my fork of sderosiaux/chrome-agent, a Rust CLI that turns web pages into structured records for agents, 42 commits in twelve hours made the test suite compile and pass on Windows for the first time (startup, Chrome discovery, process cleanup and handle inheritance fixes) and added Windows and macOS CI that fails if any test is skipped, after five had been skipping silently while reporting green. The work stays on the fork.
-
Entire: session path security report credited in two merged fixes
I reported session ID path handling defects to Entire. Its maintainers credited my original report in fixes merged into entireio/cli and entireio/external-agents in September 2026. The maintainers wrote the fixes.
Writing
I write about the systems I'm building, the work of checking them, and what happens to judgement, responsibility and attribution when more of the work is done with AI.
A few places to start
- Agent engineering and orchestrationThe Code Question grep Can't Answer
- Governance, evidence and assuranceDAG TOML: How I Turned Four Months of Code-Review Pain into a Machine-Checkable Planning Format
- Security and open-source audits3 AIs Reviewed the Same Codebase. They Disagreed on 2 Findings. That is the Point.
- Work, attribution and the economyThe Contest Was the Story. The Dust Was the Job.
- Human judgement and responsibilityThe Signature Was His. The Front Seat Was Not.
- Research and preprintsA Minimal Executable Proof for Multi-Language Contract Traceability
Browse all 120 works
120 works, grouped by subject. Cross-posts share one entry.
Agent engineering and orchestration
-
Working memory beyond grep
Commentary addressed to Matthew Berman on structural code search, call graphs and dependency impact as working memory for coding agents.
-
Making capable models useful inside organisations
Commentary on validation, consensus, provenance and execution that survives changes to the underlying models.
-
Going Remote, Without Going Reckless: Multi-LLM Orchestration and the New Front Door in llm-cli-gateway 2.9.0
Explains the gateway's remote boundary: principal isolation, registered workspaces, approval constraints and the limits of remote execution.
-
llm-cli-gateway 2.5.0: OAuth for remote MCP connectors and safer workspaces
Covers remote MCP OAuth, workspace aliases and checks that reject unsafe requests before a provider process starts.
-
Reviewing Patrick Collison's Ask for an LLM Workflow Tool
Reads Patrick Collison's workflow-tool request against public agent tooling and the longer-running studio model being built behind it.
-
State machines and deliberate human control
Commentary on XState, Stately Studio and bounded action spaces as a way to retain control around model-assisted work.
-
Routing needs a control layer
A reply on routing between models and the harness and control responsibilities that remain around local or open-weight execution.
-
A structural map of the Linux kernel with sqry
A published technical report of a sqry index over a Linux source tree, including its scale, language breakdown and graph observations.
-
Context librarians and the agent substrate
Related replies on specialised context agents, semantic search, explicit intent and independent model review around worker agents.
-
Semantic state beyond code generation
Commentary addressed to Patrick Collison on durable semantic state, coordination, evidence and governance after generation stops being the main bottleneck.
-
The infrastructure built before predictable agent work
Reflects on building semantic search, multi-model orchestration and assurance infrastructure to improve the predictability of agent-driven development.
-
llm-cli-gateway 2.0.0: the quiet supply-chain release that matters
The 2.0.0 release replaces better-sqlite3 with Node's built-in SQLite, with migration tests and explicit runtime requirements.
-
The Machine That Builds the Machine, and the Studio That Runs Itself: Two Ways to Organise an Agent Swarm
Compares Bardo's project-shaped agent factory with atelier-studio's institution-shaped orchestration and their convergent safety mechanisms.
-
A supervisor budget is not the whole agent budget
An account of token accounting that covered a supervising agent while omitting the work delegated beneath it.
-
Tracking Five Upstreams, Fuzzing the Parsers, and a Front Door: What Changed in llm-cli-gateway
Tracks vendor CLI contracts, fuzzes untrusted-output parsers and records the release work behind llm-cli-gateway's public front door.
-
The human bottleneck in parallel agent work
A reply to Addy Osmani on orchestration costs, serial merges and the human judgement that parallel agents still depend on.
-
Skill distillation and institutional memory
A reply to Tomasz Tunguz on inspectable, versioned skills and the institutional memory around an agent.
-
The interface shift from IDE plugins to agent fleets
Related replies to Simon Willison on CLI and desktop agents, the cost of verification and the infrastructure needed for reviewable agent work.
-
When text diffs flatten a graph of intent
A reply to Theo Browne on clean merges that disagree at runtime and the limits of files as the main coordination surface for agents.
-
Cache-Aware Spawning: What Changed in llm-cli-gateway, a Week On
Makes prompt caching explicit across providers through structured prompt parts and cache observability without storing conversation text.
-
What's new in llm-cli-gateway
Covers Codex session resume, Grok as a fourth provider and durable asynchronous results that survive the polling window.
-
the next software stack needs more than code generation
Argues that intent, coordination and durable semantic state become the bottlenecks once code generation gets cheap.
-
Multiple Agents, Multiple Workstreams, and the Parts That Still Break
Describes the shift from generation to coordination, using concurrent repository work and an upstream compatibility repair in grokrs.
-
Outcomes beyond the first generated result
Argues that agent work should be judged by the outcome and its continued improvement, rather than by one-shot generation.
-
How We Built a Safety-First Rust Agent CLI in Two Days Without Letting the Codebase Turn to Mush
Explains how crate boundaries, specifications, implementation DAGs and traceable evidence kept a concentrated grokrs build reviewable.
-
Decomposing work for local models
Commentary on using frontier models to decompose larger tasks into work that local models can handle.
-
Why CLI Wrapping Beats API Proxying for Multi-LLM Development
Compares CLI wrapping with API proxying for real file access, test execution and native session continuity, including the trade-offs.
-
Delegating across coding CLIs
Commentary on a shared MCP interface for coding CLIs, session management and independent models reviewing one another.
-
Token accounting and agent budgets
Published commentary on repeated context, cache accounting, agent iteration budgets and visibility into accumulated usage.
-
Structural code search for coding agents
Explains AST nodes, typed relationships and deterministic graph queries as a different working surface from token search.
-
The Code Question grep Can't Answer
Introduces sqry, which parses code into one cross-language graph so an AI assistant can answer who calls what, and compares it with grep, embedding search, ast-grep, CodeQL and Sourcegraph.
Governance, evidence and assurance
-
A content hash tells you a skill changed. It doesn't tell you the skill escalated.
Follows a skill pack into governed distribution and asks what detects expanded authority when an integrity hash only detects changed bytes.
-
Compilation Is Not Assurance: What Agentic Rust Needs After the First Green Build
What AI-authored Rust needs after the first green build: explicit ownership, concurrency, unsafe and panic claims, bounded implementation DAGs, independent validation and cost records.
-
Governing AI-assisted open-source contributions
A five-post introduction to tiered review paths, research summaries, DAG-TOML plans and a reusable contribution-governance pack.
-
Assurance when agents enter company operations
A six-post discussion of agent participation in company operations, the limits of text diffs and the independent evaluation and decision records needed around them.
-
Permissions across agent runtimes
Commentary on the permission and governance boundary when several agent runtimes can be invoked through a unified interface.
-
Token billing and fitness at the point of consumption
A six-post thread on token-billing incentives, verification costs and making fitness and evidence visible where model output is used.
-
Structural memory for agent-driven changes
A reply to Jaya Gupta on intent, dependencies, evidence and provenance as records that travel with a code change.
-
OpenRouter Fusion and consensus as governance
A six-post discussion of OpenRouter Fusion, model diversity, synthesis and the distinction between benchmark performance and a governance process.
-
The distance between a viral demo and governed operations
Commentary on validation, governance and durable semantic state as requirements beyond a successful agent prototype.
-
Diagrams and machine-checkable intent
Distinguishes the help a diagram provides from dependencies, contracts and evidence that a validator can check.
-
Artifact proofs and their evidence trail
Distinguishes Zero artifact proofs from the DAG-TOML evidence trail around them and links a small public integration demo.
-
Scaffolding that makes agent work durable
A reply to David Fowler on the foundations needed when agents take on real work and code is only one part of the record.
-
Intent and attribution across parallel agents
A reply to Peter Yang asking how intent and attribution survive parallel agents and repeated changes to a shared work product.
-
Making prompts reviewable
A reply to Drew Breunig introducing DAG-TOML as a way to record intent, dependencies, evidence and readiness for review.
-
Nondeterminism is not the same as unreliability
Distinguishes probabilistic output from the contracts, evidence, replay and review gates needed when it affects system invariants.
-
Comparing skill registries and their governance models
A comparison of dgov, Bardo and DAG-TOML, with attention to file claims, preflight checks, failure memory and fleet-wide policy.
-
Hosting skills is not enforcing them
A reply to Matthew Berman on skills as versioned contracts and the enforcement needed when an agent can skip loading them.
-
Provenance and reputation
A reply to Robin on git provenance chains and the separate role of a reputation layer.
-
A review cycle built around specifications
A reply describing a workflow of structured specifications, independent model reviews and implementation checked against the agreed plan.
-
The City-State and the Federation: Two Governance Models for AI Coding Agents
Compares dgov's agent constitution with DAG-TOML's evidence and validation model, including the blind spots each approach leaves.
-
DAG TOML: How I Turned Four Months of Code-Review Pain into a Machine-Checkable Planning Format
Mines a review archive of about 2,400 documents into six failure categories and introduces DAG-TOML, a planning format whose validator recomputes every derived claim.
-
Two formats and six classes of review failure
Related posts on separating narrative from machine-checkable structure and on the evidence, contract and boundary failures that prompted that approach.
-
The Format I Make Agents Author In: A Working Guide to DAG-TOML
A working guide to DAG-TOML: narrative and structure, document kinds, machine-checkable invariants and the review process around them.
-
Reviewing a release against source and a running binary
A published account of wiring the Grok CLI surface into llm-cli-gateway and using independent code, binary and mutation checks to review the change.
-
Skills as enforceable contracts
A reply to Daniel Jeffries on separating a skill operating manual from inherited, versioned rules and enforceable contracts.
-
Writing down the boundary before the code arrives
Argues that requirements, abstraction boundaries and required evidence must be explicit before an agent makes the change.
-
A change can execute cleanly and still be wrong
Distinguishes validated execution from a change respecting the conventions, intent and boundaries it was meant to preserve.
-
State machines for reviewable agent plans
Connects explicit UI state transitions with plans that declare their intended transitions before agents write code.
-
Discovery depends on a truthful surface
Examines how agents discover and use features from their descriptions, and the matching risk when that description is wrong.
-
Documentation freshness and the change that caused it
Examines the review and upkeep of generated documentation, arguing that intent should travel with the change that can make it stale.
-
Reviewability at the scale of generated PRs
Argues that a change must arrive with its requirements, boundaries and evidence if human review is to remain useful at high volume.
-
Model drift and layered skills
A reply to Matthew Berman on changing vendor harnesses, layered skills and the use of narrative, structure and independent review to govern them.
-
Here's what stopped breaking, when you make LLM agents author in two formats
Separates narrative, machine-checkable TOML structure and evidence so an agent's account of its work can be checked.
-
Markdown guidance in a growing context
A reply to Robert C. Martin arguing that informal Markdown guidance weakens as the surrounding agent context grows.
-
Token spend and review rigour
Reflects on repeated specification and implementation reviews, their token cost and the standards used to judge AI-generated software.
-
Moving agent plans from prose to checked structure
Commentary on a progression from Markdown and JSON to DAG-TOML and database-backed records for keeping agents on track.
-
Architecture and responsibility beyond a personal prototype
Related replies on the coordination required for useful AI-authored software and the greater safety and stability obligations when others rely on it.
-
The Codex Review Gate: How We Made AI Agents Review Each Other's Work
The review pattern used most at Verivus Labs: work goes to Codex, Gemini and Claude through llm-cli-gateway and iterates until they approve it without conditions, extended into consensus and red-team rounds.
-
How to Set Up Multi-LLM Code Review with Claude, Codex, and Gemini
A practical guide to multi-model code review through llm-cli-gateway, using sqry to give reviewers structural code context.
Security and open-source audits
-
Treating model weights as untrusted binaries
A reply arguing that open weights are untrusted inputs and that access to a model file is different from being able to audit its training.
-
Hardening API Scan Boundaries in skill-scanner, with sqry as the Review Map
Uses sqry to map the reachable API boundary of skill-scanner, then connects the hardening work to fixes, tests and documentation.
-
Skill governance inside the enterprise
A reply to Arun Sasi on internal threats, drift and accountability for judgement executed by agents.
-
Allowed actions can still be wrong
Distinguishes permission checks at the tool boundary from the separate task of governing the meaning of an allowed change.
-
Governance before agent hype
A reply to Zach Wilson on prompt injection, secrets, supply chains, identity and isolation as production concerns around agent systems.
-
Infrastructure sovereignty and weight sovereignty
Commentary distinguishing the location of model infrastructure from the provenance and trust questions around the weights it runs.
-
3 AIs Reviewed the Same Codebase. They Disagreed on 2 Findings. That is the Point.
A Claude, Codex and Gemini review of Simon Willison's llm CLI using sqry and llm-cli-gateway. Claude adjudicated 11 candidate findings (8 confirmed, 2 disputed, 1 uncertain). Three findings were filed as #1396, #1397 and #1398, and the article says none have been confirmed by the maintainer yet.
-
How We Used AI Agents to Security-Audit an Open Source Project
A two-round security audit of Garry Tan's gstack using sqry and llm-cli-gateway. Round 1 reported 10 unique security findings. The round 2 findings table lists a high-severity shell injection via bash-to-JS interpolation.
-
The harness as a security boundary
A reply to Sebastian Raschka examining the security choices in the harness around a coding model.
-
Provenance, SBOMs and install-time scripts
A reply to Gergely Orosz proposing release provenance, software inventories, constrained installs and dependency checks.
Technology in practice
-
Model costs and provenance
A reply to Peter Diamandis arguing that model price and benchmark comparisons should also account for provenance and how capabilities were acquired.
-
What Bunnings' AI chatbot reveals about how retail AI actually works
Examines Bunnings' Buddy chatbot as a concrete example of the architecture and commercial choices behind retail AI.
Work, attribution and the economy
-
A coherent strategy still needs evidence
A reply to a16z distinguishing a strategic narrative from the decisions, resource allocation and outcomes that test it.
-
Cost per PR and the outcome a change delivers
A reply to Ben Holmes distinguishing cheaper delivery from product value, operating costs and customer outcomes.
-
Pricing the judgement behind the labour bill
Commentary on labour-market valuations of AI and the missing economic claim for the people whose judgement is being reused.
-
The Contest Was the Story. The Dust Was the Job.
Uses John Henry and the symphony question to examine the work left outside the human-machine contest, and who has standing to judge it.
-
Human capital and token capital
A repeated reply and quote-post on experience being absorbed into models and the distribution of the resulting economic asset.
-
Expertise and the ownership of its returns
Commentary on the value of domain judgement in model-assisted work and who retains the benefit as that expertise is incorporated into systems.
-
Ownership needs enforceable provenance
A reply to Daniel Miessler arguing that owning infrastructure is insufficient unless provenance and contribution claims remain enforceable.
-
Human formation and a stake in recorded judgement
A reply to Jonathan Pageau adding the question of economic claims to a discussion of human formation and corporate knowledge capture.
-
Recorded company knowledge and human claims
Related commentary on meeting recording, AI synthesis and whether the people supplying judgement retain a durable claim over its value.
-
The Recorded Company Brain
Examines recorded organisational knowledge and the contributor's claim that has to be negotiated before the record closes.
-
Routing intelligence and preserving contribution claims
Two related posts connecting model routing with the economic record of human contribution and the claims that should survive it.
-
A machine beneath an economic claim
Commentary on the evidence required beneath economic claims when interviews, reputation and institutional trust are insufficient.
-
The Machine Under the Claim
Part two of the outcome-unit trilogy sets out the machinery, minimum floor and failure modes beneath durable contribution claims.
-
The Outcome Is the Unit
Part one of a trilogy arguing that the economic unit is the attributable outcome.
-
The Record With No Keeper
Part three of the outcome-unit trilogy examines settlement and an attribution record that no single keeper controls.
-
PR counts and the price of judgement
A reply to Gergely Orosz on cheap generation, expensive evaluation and the incentives created when PR counts stand in for value.
-
Empowerment and a durable claim on what people build
Argues that recognising scarce human talent should include a lasting claim on the value it creates, beyond the wage attached to a role.
-
Why abundant output can raise demand for expertise
A reply to Dan Shipper connecting automation with increased demand for human framing, verification and judgement.
-
Attaching cost to declared intent
Related replies argue for binding cost to the requirement that justified the work, rather than reconstructing value from spending dashboards later.
-
The work that grows when agents write more
A reply to Aaron Levie on setting the frame, reviewing output and keeping that review legible as agent output increases.
-
New work beyond productivity gains
A reply to Francois Chollet on AI enabling new workflows while discernment and the choice of problems remain human responsibilities.
-
Automation and the human work that expands around it
A reply on growing human work around automated execution, with framing, verification and governance becoming more important.
-
Cheap generation and the responsibilities it leaves behind
Connects open-source supply-chain responsibility with the gap between the cost of starting thought and the cost of judging its value.
Human judgement and responsibility
-
Discernment needs standing and recourse
A ten-post thread on the authority, visibility and recourse people need to challenge machine-framed decisions and change the conditions that produce them.
-
The Signature Was His. The Front Seat Was Not.
Distinguishes naming a human to carry responsibility from giving that person the visibility and authority needed to exercise it.
-
Human intent and its delivery layer
A reply to Trevin Chow distinguishing human intent from the machinery that carries and realises it.
-
The Provenance Was Valid. The Meaning Was Compromised.
Reads the Mini Shai-Hulud npm worm (valid SLSA provenance from a hijacked CI pipeline) through Kahneman, Duke and Schulz to argue that a valid process does not guarantee a valid meaning.
-
Learning with papers and multiple model perspectives
Commentary on using research papers and several model providers to explore a subject and develop durable understanding.
-
Craft as the tools change
Commentary on discernment, architecture and judgement remaining part of the craft as coding tools become more powerful.
-
Buying execution and outsourcing judgement
A reply to Peter Diamandis on the make-or-buy decision moving upstream from execution into deciding what work is worth doing.
-
Who Dealt the Final Card?
Reads decision quality and uncertain outcomes through Annie Duke, Daniel Kahneman, Kathryn Schulz and a Reformed account of providence.
-
Holding the standard when work cannot be reconstructed by hand
Commentary on leadership as maintaining the standard by which generated work can still be governed and checked.
-
Keeping discernment while buying cheap execution
A reply to Peter Diamandis on better questions, the make-or-buy decision and the danger of outsourcing the ability to evaluate answers.
-
Keeping the verification loop in AI-assisted accounting
Commentary on buying cheap execution while professionals retain the ability to distinguish useful results from fabricated ones.
-
What a truth system refuses to assert
A short reflection on refusal as a human and moral responsibility within systems that make claims.
-
The make or buy decision comes for your mind
Applies Ronald Coase's make-or-buy question to thought: asking became cheap, checking stayed expensive, and judgement still has to be governed.
-
Syntax, systems thinking and the reasons beneath them
Commentary distinguishing tool syntax from architecture and systems thinking, then from the philosophical questions beneath both.
Research and preprints
-
A Minimal Executable Proof for Multi-Language Contract Traceability
An arXiv preprint connecting a small multi-language executable proof to contracts, a DAG, traceability and evidence. It reports Java as skipped and makes no production-readiness or general semantic-equivalence claim.
Experience
I have spent over 25 years in IT, starting as IT Infrastructure Manager at Internet Solutions in South Africa in 1999. Since then I have been a Senior Systems Architect at Unisys, Account CTO at Microsoft Australia, Director of IT Services at Thomas Hassall Anglican College, team lead for 7 to 14 solution architects at the City of Gold Coast and Senior Staff Systems Architect at ServiceNow, to list a few.
-
Founder, Verivus
Current work is a governed platform for business decisions and actions, customer-controlled access from AI agents to tools and data, and developer tools for understanding code and coordinating coding agents. The Assurance Platform and hosted AgentFederator service are in development; sqry, llm-cli-gateway and agent-assurance are released as open source.
-
Senior Staff Systems Architect, Cloud Infrastructure Systems Engineering
Architecture for sovereign-cloud platforms spanning Australia, Singapore and Canada on one of the world's largest SaaS platforms, including a CIS Controls v8 compliance program; recognised with two UTG Llama Awards.
Earlier experience (9)
-
Principal Enterprise Architect (Freelance)
Created the QBCC Digital Services Reference Architecture and EA repository (incorporating the QLD Digital Economy Strategy), and delivered a ServiceNow future-state analysis and roadmap for the QLD Department of Education.
-
Citrix Business Development Manager (contract)
Built the Fujitsu-Citrix partnership and the TTU @ Fujitsu sales motion, growing weighted first-year order-value pipeline from AU$484K (May) to AU$5.011M (November 2022).
-
Enterprise Solutions Architect / Team Lead, Solution Architecture
Led a team of 7 to 14 solution architects delivering citywide ICT programs on Azure and AWS for Australia's second-largest local government (3,200+ staff); delivered the City's first Virtual Council Meeting in two weeks at the onset of COVID.
-
Principal Systems Administrator, Data Centre & Cloud
Technical leadership for operations across Azure, AWS, VMware (vSphere and Horizon), enterprise servers and SAN supporting Queensland's police, fire and emergency services; technical lead for the data protection and recovery tender and HP OneView rollout.
-
IT Infrastructure Team Lead
End-to-end infrastructure for a utility serving 40,000 households: vSphere 6.0 and Active Directory upgrades, ADFS, SharePoint build for EPM, and technical lead for the EMC Unity all-flash array implementation.
-
Director of IT Services
Brought IT in-house from full outsourcing for a K-12 school of 1,650 students and 170 staff; built a Microsoft private cloud (System Centre + Hyper-V), deployed Office 365 for 1,650 students, and modernised campus networking and AV.
-
Virtualisation Technology Specialist; Lead Technology Strategist (Account CTO); Sales Excellence Program Manager
Six years across three roles: led the Hyper-V Rapid Deployment Program to 200% of its design-win goal against VMware, served as Account CTO exceeding quota four consecutive years, and built a BI platform surfacing more than $1.8B in whitespace revenue opportunity.
-
Senior Systems Architect
Lead architect on enterprise consolidation and infrastructure programs delivered on time and within budget for Westpac, the ATO, NSW Roads and Traffic Authority, South African Reserve Bank, Nedbank, Telkom, Old Mutual, and Konkola Copper Mines.
-
IT Infrastructure Manager
Infrastructure management for South Africa's largest corporate ISP.
Awards
-
UTG Llama Award, Q1 2025
Awarded for resolving regulated-market customer issues in Australia and taking on hyperscaler program work beyond my own responsibilities.
-
UTG Llama Award, Q2 2024
Awarded for preparing and reviewing government customer engagement material that became the basis for ServiceNow's isolated-cloud stakeholder communication.
-
Oceania Hub Award
Awarded for the TTU @ Fujitsu mini-site supporting the Fujitsu-Citrix partnership sales motion.
-
Delivering Results Australian Subsidiary Award
Microsoft Australia subsidiary recognition received during the Lead Technology Strategist (Account CTO) tenure.
-
Gold Global Recognition Award
Received in 2005 during the Senior Systems Architect tenure covering enterprise consolidation and infrastructure programs across Australia and Southern Africa.
-
Worldwide Architects Forum
Appointed to the Unisys Worldwide Architects Forum in 2004.
Education
-
Master of Networking and Systems Administration
Conferred by Charles Sturt University on 10 April 2008.