Forward Deployed Engineering · JAPAC · Google Cloud · Melbourne, Australia
I work on the layer between a model that can do something and a system you can trust to keep doing it: compilers, runtimes, control planes, durable execution, context containment, synthetic worlds, evaluation infrastructure and developer tooling.
The model can improvise. The system should know what happened.
| NOW | straitjacket — context economics for coding agents |
| RECENT | Anvil — legacy estates → agent-safe capability surfaces |
| BUILDING | Worldloom — coherent synthetic enterprises and eval truth |
| If your problem is… | Start with |
|---|---|
| Agents need to use APIs or legacy middleware safely | Anvil |
| A coding agent is drowning in tool output | straitjacket |
| An acting agent must survive crashes and partial effects | Tape |
| You need realistic enterprise corpora with ground-truth evals | Worldloom |
| You need a realtime Gemini voice/runtime layer in Rust | gemini-rs |
| You need to industrialize intent → agent → proof → admission | GE Agent Factory |
Worldloom · gemini-rs · GE Agent Factory
INTENT
│
├── contracts / specs ───────► Anvil ─────────────► CLI · MCP · skills · hooks
│ └───────────────► legacy API estates
├── agent design ─────────────► GE Agent Factory ─► code · evals · passports
├── runtime semantics ────────► gemini-rs ────────► voice · state · governed flows
│ └───► Tape ─────────────► journal · replay · effects
├── fleet / policy ───────────► AIPlex / Scion ───► identity · isolation · routing
├── context pressure ─────────► straitjacket ─────► bounded digests · exact retrieval
├── synthetic reality ────────► Worldloom ────────► facts · artifacts · eval truth
└── human surfaces ───────────► Pixelpitch / ge-msft
More systems — control planes, DX, multi-agent execution and human surfaces
| System | What it explores |
|---|---|
| AIPlex | One policy plane across agent↔tool, agent↔agent and agent↔model interactions. |
| adk-fluent | Python + TypeScript fluent builders generated from a shared manifest into native ADK objects. |
| Scion | Isolated container/worktree execution for collaborating deep-agent harnesses. |
| Pixelpitch | Agent-authored HTML → editable PPTX under explicit visual-fidelity constraints. |
| ge-msft | Gemini Enterprise inside Microsoft 365 with reversible, provenance-bearing actuation. |
| antigravity-a2a-a2ui | Identity-derived per-user managed-agent workspaces and credential brokering. |
| Thesis | Systems it pressures |
|---|---|
| Agents are at Docker. We think we need k8s. We actually need CNCF. — primitives create coordination problems; durable value moves up-stack. | AIPlex · Scion · GE Agent Factory |
| Winter is Coming — protocols, frameworks and infrastructure must deliberately compose. | Anvil · Tape · gemini-rs · AIPlex |
| The Asymptote of Good Enough — optimize only while marginal improvement changes behavior or outcomes. | straitjacket · adk-fluent · Worldloom |
| Tacit Code, Real Friction — as code gets cheaper, integration, verification and institutional constraints become scarcer. | Anvil · ge-msft · eval infrastructure |
This README now dogfoods the same architecture I keep reaching for elsewhere:
portfolio.yaml
│
├── README-facing project diagrams
├── social / Open Graph card
├── standalone landing page
├── generated portfolio index
└── activity surface
portfolio.yaml is the canonical model. scripts/generate_portfolio.py emits the projections. profile-assets.yml regenerates them when the model changes.
The social card is assets/og-card.svg. The site projection lives under site/. One source of truth; multiple surfaces.
01 SPEC > PROMPT durable intent should outlive a model call
02 IR > HAND-WIRING compile multiple surfaces from one source of truth
03 REFUSE > GUESS uncertainty is information
04 RESUME > RETRY acting agents need memory of reality
05 ADDRESSES > SUMMARIES omitted bytes need a deterministic path home
06 EVIDENCE > VIBES performance claims need receipts
07 PURE COMPOSES; EFFECTS GATE side effects stay explicit and reviewable
08 LOCAL FIRST cloud cost and mutation should be opt-in
09 BORING CONTROL PLANE spend stochasticity only where it buys leverage
10 DEVEX IS ARCHITECTURE setup, errors, introspection and docs are system design
The newer agent work sits on older layers: autonomous-driving projects around lane geometry, behavioral cloning and traffic-sign recognition; then Kubernetes, CI/CD, landing zones and operational plumbing. The technologies changed. The recurring interest did not: how do you turn uncertain behavior into an engineered system with explicit boundaries?
browse the code · read the thinking