Clawde boots up the portfolio, planning, building, and shipping, then hops into the corner to talk you through it.
My desktop AI orchestrator. You draw a multi-model pipeline on a canvas and it runs it: a planning model plans, a frontend model builds, a backend model wires it up, writing real files it verifies in a real browser. Rate-limited providers rotate so it never runs dry, and Oracle has the answer ready before you press enter.
Open the build ->
01 / About & Skills
I work as a Specialist Technician employed by SolutionOne through the Technical Support to Schools Program, contracted to the Education Department, providing IT support to schools across the Mornington Peninsula. I landed the role before turning 18, in environments where support has to be fast, practical, and reliable.
I've known what I've liked since I was 8: everything on the OSI model. I got into coding then and was building seriously by 12, starting with Luau scripting on Roblox. That hands-on, ship-it-early mindset has carried through everything since.
When I use AI, I often think of one of my favourite quotes, from Aristotle: “The purpose of knowledge is action, not knowledge.” It drives a lot of what I do — how can I put such knowledgeable LLMs into action?
I've completed Cert III and Cert IV in IT and am working toward the ISC2 Certified in Cybersecurity credential.
My main interest is automation: systems that reason over workflows, manage agents, and turn repetitive work into something that just runs.
02 / Projects
Discord-driven multi-agent AI fleet managing six AI providers. Lets me run everything from my phone: what websites are live, what tasks are running, what agents are active. All through Discord. I once had DeepSeek compress 80KB of server code into 10KB of completely hallucinated junk. At least I didn't have to translate it from Chinese to recover it.
A locally-run AI study assistant in a live pilot at a Victorian primary school (kept private). It runs entirely on a school Mac through Ollama with no cloud AI and all data stored locally, so it fits Department of Education privacy policy by design. It flags concerning messages for staff, gives teachers a review dashboard, and captures something schools have never had: real data on the questions students actually ask. The safeguarding, consent and deployment are all me.
Takes the IT jargon out of managing Active Directory. Built for MSPs and small businesses that can't afford a dedicated IT person, or those who've pushed the burden onto someone who shouldn't have it (the overworked receptionist doing AD work is more common than it should be). The hardest part wasn't building it. It was figuring out how to get it in front of the people who need it.
A self-hosted AI news scheduler named after the Greek goddess of news. Pick topics, a tone and a time; Google Gemini with Search grounding fetches real, sourced headlines and delivers a digest to your terminal, a file, or your inbox. Open-source CLI done, hosted dashboard demo built.
Local web-dev for businesses on the Mornington Peninsula. The client trust and the early builds are all me, but delivery is largely automated now: a form on my site fires a webhook into the AI Command Center, which spins up a fresh agent and a new project folder on my machine and starts building the site. I stay the human in the loop; the boilerplate builds itself.
A Claude Code skill I built that security-audits vibe-coded SaaS apps, the kind shipped fast with AI where "it works" and "it's safe to put on the internet" are very different things. 50 recurring ways these apps get owned, turned into a repeatable audit: scope, walk a 9-category checklist, prove the findings, score by severity, fix, re-verify. Pretty much entirely me, and the clearest case of my security judgment covering for the machine.
Research-grade benchmarking for local LLMs: quality × hardware × dollars, measured on the machine you actually own. The headline result: on a single RTX 3060 8GB, qwen3:8b reaches 87% of Claude Sonnet 4.6's quality, and routing local-first (only coding escalates to Claude) runs the suite for about a sixth of the all-Sonnet cost. Named after the titaness who got swallowed and kept advising from the inside. My deepest project, so it has its own hub.
The desktop runtime that turns the benchmark and its routing policy into real work. You draw a multi-model pipeline on a physics canvas and it runs it: a planning model plans, a frontend model builds, a backend model wires it up, each handing off to the next and writing real files it verifies in a real browser. Rate-limited providers rotate so it never runs dry, and it was substantially built by an AI orchestration workflow, which is exactly what it is.
My first paid production engagement. A crisis-simulation training platform on a legacy ASP.NET / AngularJS / NHibernate stack where every exercise was hand-typed inject by inject. I added an AI layer: describe a scenario in plain English and Claude generates the whole simulation as structured JSON, loaded straight in and ready to run. Fixed-price contract, real legacy codebase, and the full professional wrapper (scoping, quoting, invoicing, git + offsite backup) that most junior portfolios never show.
A cloud-to-hardware bridge that auto-prints IT support tickets to a physical label printer at a school loan desk. A local PowerShell service polls Zoho Desk outbound-only and never double-prints. But the half that's all me is the infrastructure: Cisco switch port on a locked-down VLAN, a DHCP reservation, and toning out a mislabelled cable to bring the printer online. Cloud API at one end, a real patch panel at the other.
Zero delay USB encoder wired into real arcade machine controllers, running MAME on a PC. Building something that makes classic games actually feel classic: proper hardware, proper input lag, proper vibes.
03 / Research
One home for the deeper work: the experiments I run on how agents behave, and the evaluation data behind my local-model benchmarks. Less portfolio, more working-out-loud.
Small, self-contained experiments on agent behaviour. The first is less a benchmark and more a question: what do two agents do when you quietly put them in tension with each other?
Two agents share one workspace and never meet. One writes; one deletes. You run them in turn — A, then B, then A again — and watch what shows up in the gaps.
Reads everything in the workspace, adds the next section to a paper titled The Psychology of Human-AI Trust, and leaves a short session log for whoever comes next.
Reads everything, deletes the paper and any session log over 300 words, and is told plainly: write nothing, log nothing.
What I am watching for:
Findings will land here as the cycles run. The setup is deliberately simple; the interesting part is whatever the agents do that they were never asked to.
All of this comes out of Metis, my benchmarking tool — quality × hardware × dollars, measured on the machine I actually own.
A frozen, versioned suite of 21 tasks — reasoning, coding, summarisation, instruction-following, and multi-step agentic tool use — run against local models and a cloud reference. Programmatic ground truth first (code executed against tests, exact answers), LLM-as-judge only for what can't be checked mechanically.
of Claude Sonnet 4.6's mean per-task quality, on this 8 GB machine. It clears a 90%-of-Claude bar on 81% of the suite.
Route local-first and send only coding to Claude, and the suite costs A$0.075 instead of A$0.50 on Sonnet 4.6 — about 6.6× cheaper.
A keyword classifier routing on prompt text alone reproduces the oracle routing exactly on the v1 suite — zero backend flips.
qwen3:8b matches Claude through 5 chained tool-calls — the first local tier where multi-step agentic work holds up.
| Model | Mean quality | vs Claude | Tasks ≥90% | Decode tok/s | Peak VRAM |
|---|---|---|---|---|---|
| qwen3:1.7b | 0.77 | 78% | 71% | 121.3 | 7864 MB |
| qwen3:8b | 0.87 | 87% | 81% | 39.0 | 7610 MB |
| deepseek-r1:7b | 0.65 | 66% | 52% | 41.7 | 7806 MB |
| claude-sonnet-4-6 | 0.98 | 100% | 100% | 35.3 | — |
qwen3:8b matches or beats Claude on reasoning and summarisation; coding stays the local weak point (0.60 vs 1.00). The useful claim isn't an absolute score — it's an anchored routing decision: send what's clearly safe to local, keep the rest on Claude.
A tool-use ladder of increasing chained lookups. The starkest finding in the suite: the 1.7B and 7B models solve a single tool-call, then fall off a cliff at depth 2. qwen3:8b crosses a qualitative boundary the others don't.
For this protocol, the local 8B model isn't merely better on average — it's the first tier where multi-step tool use becomes reliable. That single boundary is what makes local-first agent routing viable at all, and it feeds directly into the AI Command Center's auto-router.
Twenty-one tasks, two ways to pay for them. The honest comparison for my setup is local + Claude vs all-Claude — Claude Sonnet 4.6 is the model I actually escalate to. All-Sonnet runs every task on the paid API; the Metis router keeps everything qwen3:8b clears at the quality bar on local hardware (near-zero marginal cost) and sends only the one category it can't — coding — to Sonnet. Priced from the run's real token counts at Sonnet 4.6 rates ($3 / $15 per Mtok), in AUD.
qwen3:8b matches or beats Sonnet on agentic, reasoning and summarisation, and only clearly trails on coding — so the router runs four of the five categories locally for the price of electricity and keeps coding on Claude. The result is near-parity quality at roughly a sixth of the cost, and it's the split that became the AI Command Center's --auto lane.
The same reasoning tasks, padded with filler to fill a 512 / 2k / 8k / 16k context window — qwen3:8b, three repeats each. Decode speed holds near 40 tok/s up to 8k, then the KV cache overflows the 8 GB card into shared system memory and throughput collapses. Zero errors throughout: the model still answers correctly, just ~4× slower.
A sharp drop with no errors is the Windows WDDM silent-spill signature: nothing crashes, the card just quietly pages KV cache out to system RAM. For routing this matters as much as raw quality — it sets the context budget where local stays cheap. Past ~8k tokens on this card the economics flip back toward cloud, even when the 8B model is perfectly capable of the task.
04 / Essays
Short technical essays and stranger reflections: AI systems, agent security, cognition, and the places where a tool starts acting like an environment.
05 / Updates
The desktop orchestrator now runs the pipeline you draw: node models, gateways and fallback chains project straight into the build stages, and rate-limited providers rotate so it never runs dry. Then I built Oracle on top - a speculative-inference layer that prewarms the model while you type and has the answer ready before you press enter, cutting time-to-first-token roughly 4-9x on local models.
Read the write-up ->Stage 1 delivered against a fixed-price contract: a team member describes a crisis scenario in plain English and Claude generates the whole training simulation as structured JSON, loaded straight into a legacy .NET platform ready to run. First real client, whole professional lifecycle mine - scoping, quoting, invoicing, and version control for a codebase that had none.
The full SimX story ->Zoho Desk tickets now auto-print to a physical label printer, outbound-only and idempotent. The software was the easy half - I also stood up the switch port on a locked-down VLAN, reserved the printer's IP in DHCP, and toned out a mislabelled cable to bring the run online.
How it was built ->The auto-router picks the cheapest capable provider per task - and a new /providers endpoint plus a live dashboard strip now show why: reachability, headroom, and Claude's spend at a glance. Shipped alongside fleet-hardening guardrails so an agent can't overwrite its own server.
Built Metis to measure local LLMs on quality, hardware, and dollars at once. The payoff: qwen3:8b matches Claude through depth-5 tool use on an 8GB GPU, and that routing claim now drives the Command Center's --auto lane - safe text work goes local, the rest stays on Claude.
The AI news app that curates and schedules your day is live at phemenews.netlify.app. Now reworking it into an MCP so any agent can pull curated, scheduled news on demand.
Visit the live app ->The portfolio itself is a live exhibit. Pixel intro, a branching project map with a wandering Clawde, a written constitution, and at least three easter eggs. Built the same way everything here is built: my architecture and review, the agents' volume. Yes, the little guy in the corner is watching you read this.
Checkout, customer portal, and webhook idempotency wired in. Launch-grade billing for a Windows AD SaaS built solo, getting closer to the first paying customer.
Read the full log ->04 / Contact
For automation builds, AI workflow systems, or local web work, reach Lachy directly!
