Independent AI systems architect · programmer · operator
North Queensland, Australia
I build autonomous systems where models, tools, infrastructure, hardware, safety and operations have to agree.
My work spans local-first agent infrastructure, resilient communications, distributed control planes, model evaluation and optimisation, RAG, multimodal generation, embodied AI and evidence-gated field systems. I do not count scaffolds as finished systems and I do not report simulated passes as execution evidence.
- Sovereign agent infrastructure: Hermes Agent, OpenClaw and Paperclip systems running as independently supervised services.
- Resilient communications: Reticulum/LXMF transport with a fail-closed optional Hermes policy gateway.
- Distributed agent operations: authenticated fleet enrolment, durable task dispatch, replayable events and bounded workers.
- FR3K infrastructure: community-owned communications research for North Queensland, controlled by explicit regulatory, safety and evidence gates.
- Local model engineering: 88-model Ollama inventory, quantisation, context/VRAM sweeps and task-specific evaluation across an RTX 3060 and constrained GPUs.
- Embodied and multimodal AI: Reachy Mini physical I/O, speech/vision/action routing, ComfyUI, FLUX, Wan and local video pipelines.
- Agent evaluation: golden datasets, drift detection, trajectory replay, cost-aware routing and local/cloud comparative harnesses.
| Repository | What it demonstrates | Executed evidence |
|---|---|---|
| hermes-reticulum-platform | Independent Reticulum/LXMF transport plus a bounded Hermes policy gateway | 38 tests; signed virtual delivery; gateway E2E; restart persistence; transport survives gateway failure |
| hermes-fleet-control | Secure local-first control plane for distributed Hermes nodes | 16 tests; HMAC heartbeats; durable task queue; authenticated events and metrics |
| fr3k-infrastructure | Executable governance and evidence gates for community communications infrastructure | Pack checksum verification; automated G0 decision gate; explicit blocked-state handling |
| qwen36-bench | Reproducible Ollama optimisation on constrained hardware | 35B MoE model on 4 GB VRAM; 5.71 tok/s peak; all short-suite checks passed |
| trajectory-replay-demo | Record and replay agent tool trajectories for model/prompt A/B testing | SQLite-backed replay and comparative evaluation |
| agent-cost-router-demo | Route tasks to the cheapest viable model instead of defaulting to frontier inference | Measured cost-routing baseline |
| mcp-rag-pgvector-demo | Production-shaped RAG with pgvector, embeddings, MMR and evaluation | Reproducible pipeline and evaluation harness |
| ollama-local-rag-demo | Fully local RAG with Ollama and FAISS | Zero-cloud retrieval and generation path |
| agent-eval-harness-demo | Golden-dataset agent evaluation and drift detection | 25-question baseline with local and hosted backends |
| divine-arsenal | Domain-specific computation joining ephemeris data, planetary hours and operational rules | Deterministic Python implementation rather than symbolic prose |
| openclaw-workshop | Two-day hands-on curriculum for agents, MCP and AI coding tools | Structured labs and deployment material |
- Operational truth outranks mythology. Read the running system, not just its design document.
- Transport and control are separate failure domains. Optional AI must not weaken the underlying system.
- Evidence is graded. Observation, unit tests, integration tests and live acceptance are not interchangeable.
- Security boundaries are explicit. Inputs are untrusted, queues are bounded, privileged audit fails closed and cloud egress is gated.
- Hardware claims require hardware. Virtual/LAN acceptance is never relabelled as RF, robotics or field acceptance.
- Useful failure reports beat polished fiction. A blocked gate is a valid engineering result.
- governed FR3K bench and field architecture for Magnetic Island and North Queensland
- Reticulum/LXMF identity, routing and offline messaging systems
- a distributed Hermes fleet with durable orchestration and independent node operation
- local multimodal and embodied-agent pipelines with measurable latency, memory and reliability constraints
- benchmark-gated self-improvement loops for agents, prompts, skills and local models
- private trading and operational systems where risk controls remain separate from model output
If you want the fastest high-signal inspection:
- Read the engineering dossier.
- Inspect the acceptance matrix in hermes-reticulum-platform.
- Inspect the security boundaries and tests in hermes-fleet-control.
- Inspect the blocked-state discipline in fr3k-infrastructure.
- Reproduce the measured sweep in qwen36-bench.
Contact: fr3k@mcpintelligence.com.au