Zac Page · AI Systems Architect · Sydney
Design it.
Build it.
Ship it.
AI that works with the internet unplugged, can’t make things up, and never sends your data anywhere. Private tools you own outright — no cloud, no subscription, no account. Buy one today, or commission a custom build for your workflow.
✓ Works offline
✓ Can’t hallucinate
✓ Your data never leaves
✓ No subscription
10+
Products live
14
Systems designed
75
VM opcodes
$9–$499
Price range
Browse products
Mac apps, offline tools, PDFs · $9–$79
Hire Zac
Custom build, prototype, or workflow audit · from $499
The creative path · built from first principles
Private AI, powered by deep engineering.
Vertexpace is not a wrapper. It is a full-stack AI ecosystem designed to run offline on your hardware, giving you control over your data, memory, and decisions. Below is the technical architecture that makes this possible.
01
Language
Light Code
75-concept semantic grid. Every answer traces to a fact you wrote. Built to keep answers tied to facts you control.
Built when I needed a language that could run inside a VM at 35 KB.
02
Runtime · AI-OS
Lumina OS
Full AI operating system: VM, agent scheduler, IPC bus, marketplace, and a CONSENT gate on every dangerous op. Runs on Mac and Pi.
Built when no existing OS gave agents real memory, real state, and real accountability.
03
Language
NovaGlyph
Integration language across all computational planes. PREDICT, DECIDE, CREATE, and LEARN — all operating on the same 64 bytes with different meanings.
Built when Light Code needed a layer that could reason across modes, not just execute.
04
Computation
POLYGRID Planes
Five planes (PREDICT / DECIDE / CREATE / LEARN / CONVERGE) that reuse the same 64 bytes with different semantics. No runtime bloat — same memory, new meaning.
Built to solve the problem of adding intelligence without adding binary size.
05
Hardware
Lumina STONE
ESP32-S3 firmware with a 64-LED concept grid, haptics, and a full LuminaVM in 18 KB of code. AI in your hand, no cloud required.
Built to prove sovereign AI works at the hardware layer, not just software.
06
Cryptography
Rubik's Crypto
Argon2id + AES-256-GCM + X25519 ECDH private messaging. Passphrase-seeded permutation of the concept table as the semantic obfuscation layer.
Built because private messaging between agents needed math, not trust.
07
Sovereignty
CONSENT Wallet
Single-dial gate on every dangerous OS operation. Ed25519 signed, hash-chain logged, LightNet v2 packet interlock. You approve it or it does not happen.
Built because every AI system needs a veto the user actually controls.
08
Network
LightNet
P2P encrypted mesh chat using Curve25519 + AES-GCM. Purple-query mode reaches peer's Baby AI. Runs on Mac and iPhone — no server in the middle.
Built because agents that cannot talk to each other are not a system.
09
Kernel · Liveness
Genesis Grid Eviction
Every agent starts with a liveness budget of 100. Each fault batch costs −25. Each successful IPC send earns +10 back. Budget reaches zero: the agent is inserted into evicted_agents — a HashSet — and is never respawned or queued for dream-cycle rewrite. A separate path handles silent zombies: any agent that runs for 500 consecutive ticks without sending a single IPC message is also permanently evicted, regardless of budget.
Think of it like a security deposit. Agents put 100 points down when they start. Bad behaviour costs them. Talking to other agents earns points back. Hit zero and you are out — permanently. The kernel does not retry broken things forever.
Built because a runtime that keeps restarting failing agents indefinitely is hiding the problem, not solving it.
10
VM · Biological
TritVM + Plasticity Decay
A secondary VM running on trit values (−1, 0, 1) mapped to the 8×8 concept grid. Every 7 execution steps, a pseudo-randomly chosen memory cell drifts back to zero unless actively reinforced — a Hebbian forgetting model inside a Rust VM.
Built to experiment with biological memory dynamics in a deterministic runtime.
11
OS · Autonomous
Dream Cycle
The OS has a dream cycle. When idle, it reviews faulted agents and attempts autonomous rewrites. Triggerable via HTTP at POST /dream/trigger. Requires dual-approval — both user and AI must vote before it runs.
Built because an OS that only acts when prompted is not autonomous.
12
Sovereignty · Mutual
Dual-Approval Consent
Sensitive kernel actions — agent rewrites, spawns, kills, memory writes, dream cycles — go through a ConsentProposal with two independent votes: user_vote and ai_vote. The proposing party auto-approves their own side; the other side starts as Pending. both_approved() must return true before execute_proposal() runs the action. Fair warning: in the current implementation, the AI's vote on user-proposed actions still requires a human to confirm that side — the framework for autonomous AI voting is in place, but full AI autonomy on approvals is an open design question.
The goal is that neither side can force a dangerous operation unilaterally. Even if that is not fully resolved yet, designing the system around mutual approval from the start changes what gets built — you cannot add consent as an afterthought.
Built because a system where only one party holds the keys is not a partnership.
13
Language · Design
Emotional Word Mapping
The 75-word instruction set maps urgency and time directly to arithmetic. urgent = ADD (10), now = SUB (11), soon = MUL (12), later = DIV (13). alive = CMP (30), dead = JZ (31). The section in lang.rs is commented "Urgency / ALU ops." This is not metaphor — these are the literal concept IDs the VM executes.
For bigger systems this has practical uses beyond aesthetics. In a mesh where programs transmit as bytes over LoRa or P2P, the urgency of an operation is encoded in the instruction itself — no metadata needed. In audit logs, what an agent was doing at any tick is human-readable without a decoder. In multi-agent coordination, an agent broadcasting urgent carries a different semantic weight than one broadcasting later — and other agents can respond to that difference without an extra protocol layer.
Built because a language you write by hand should feel like something, and because encoding intent in the instruction reduces the need to explain it separately.
14
Host · Execution
Cortex + Motor Cortex
The Cortex Hub (port 8889) has two layers. Cortex: thinking — content generation, AI variations, prompt refinement via Ollama. Motor Cortex: action — natural language converted to shell commands by Ollama, then executed on the real OS. Thinking and moving, separated.
Built to mirror the brain's own architecture: intention is not the same as movement.
Available now · one-time purchase · no subscription
The catalog.
Mac App · Free
Prompt Architect
Deterministic AI auditor. 12 failure-logic vectors, 8-dimension scorer, LLM-as-Judge pipeline. Catches sycophancy, hedging, and instruction drift before they cost you.
Free
Download free Mac App · $29
Fix AI
Offline computer diagnostic for Mac and Windows. Reads disk, memory, crashes, and network before you type — then diagnoses and fixes.
$29
Buy now Mac App · $29
Baby AI Lite
Private AI chat that runs entirely on your Mac via Ollama. No cloud, no account, no telemetry. Unplug Wi-Fi mid-chat and it keeps working.
$29
Buy now Mac App · $49
Lumina OS
A private AI-OS runtime. Write agents in Light Code, run them in a round-robin VM scheduler, and route every dangerous op through a CONSENT gate.
$49
Buy now Mac App · $79
NovaGlyph Kit
34 ready-made AI templates you edit, recompile, and ship. ~35 KB native binary. No cloud, no API key. Every answer traces to a fact you wrote.
$79
Buy now Web App · $19 each
Offline Specialist Tools
Domain-specific offline HTML tools: Pool Pro AI, Homelab Pro, First Aid Pro, Beekeeping Pro, Dog Care Pro. 35–70 expert answers each. Any browser, works without internet.
$19 each
Buy now PDF · $9
Light Code Cheatsheet
A whole programming language on one page — 64 concepts on an 8×8 grid plus extension rows. Print it, pin it, ship faster.
$9
Buy now What I build for you · cloud or local
The work I take on.
How I approach client work
Small, sharp AI systems that remember and act.
My strongest client work pairs an LLM with a symbolic layer: the LLM handles conversation and rewriting, while the symbolic layer stores facts, plans, domain state, and learned outcomes. The product becomes intelligent through what it does — not just what it says. It runs locally, keeps knowledge on your device, and makes useful decisions without a data centre behind it.
Who this is for
Founders, solo operators, creators, trades, educators, and small teams who need a working AI tool built around their real process — not a generic wrapper.
Where I work best
Messy workflows, private data, offline environments, specialist knowledge, repeated decisions, diagnostics, and tools that must keep working without a cloud subscription.
What makes this different
I do not start with a giant model and hope. I design a small system — memory, symbols, rules, state — then use the LLM where language genuinely helps.
AI Systems Architecture
I design the full stack: memory model, agent loop, symbolic state, LLM interface. The intelligence lives in how the pieces connect, not inside any single model.
Language & Compiler Design
I build the language layer when existing tools are not enough. Light Code, NovaGlyph, POLYGRID planes — working grammars, assemblers, and native compilers.
VM & Kernel Engineering
Full VM interpreter, round-robin scheduler, IPC bus, agent marketplace, bare-metal Pi kernel. I write the runtime, not just the app that runs on it.
Small Offline Intelligence
Products that work without cloud or data centres. A compact local engine handles decisions, memory, and useful actions — on Mac, iOS, ESP32, or Pi.
Product Build & Launch
I take the idea from concept to packaged tool: UI, local runtime, Stripe checkout, Electron bundle, macOS app, download tokens, and a supportable product page.
Hardware & Firmware
ESP32-S3 firmware, bare-metal Pi 5 kernel, Swift iOS VM port. I build the software that runs on the silicon, not just the service that talks to it.
Proof from the work
What I've actually built.
Rust · VM + Kernel
Lumina OS
A full AI operating system — scheduler, IPC, agents, CONSENT gate.
75-opcode interpreter, 64 KB RAM per agent, round-robin scheduler, agent marketplace, self-modifying supervisor. macOS and Pi.
Rust · Native ARM64 Compiler
LightNet / lightc
A compiler and P2P mesh chat, both written from scratch.
Compiles the 75-concept Light Code language to native ARM64. Includes a working REPL and E2E encrypted mesh chat (Curve25519 + AES-GCM) on macOS and iOS.
Rust · Electron · macOS + Windows
Baby AI Lite & Fix AI
Two packaged Mac apps, sold via Stripe, running fully offline.
Baby AI Lite: local Ollama chat with SQLite memory. Fix AI: reads real disk, memory, and crash state before responding. Both signed and live.
Rust · Compiler · ~35 KB binary
NovaGlyph Kit
A compiler for fact-bound specialist tools.
34 templates. Write a facts file, run one command, get a 35 KB native binary. Every answer traces to a fact you wrote — no model required at runtime.
Next.js · Stripe · Vercel
This storefront
Checkout, tokens, and fulfilment — solo built and maintained.
Stripe checkout, encrypted download tokens, webhook fulfilment, multi-platform download detection. Everything on this site.
Rust · ESP32-S3 · Swift · Pi
Hardware family
Three devices running real code today.
STONE: ESP32-S3, 64-LED grid, haptics, 18 KB firmware. Bare-metal Pi 5 kernel (no Linux). Swift iOS LuminaVM port — 6/6 tests green.
TypeScript · OSS
glyphshrink
Compresses Claude prompts without a wrapper service.
Drop-in Anthropic SDK wrapper using the 64-concept grid. 13–25% reduction on prompts, up to 70% on multi-turn agent loops.
Rust · LightNet
CONSENT wallet
A consent gate that defaults to no.
Single-dial gate on every dangerous OS op. Ed25519 signed, hash-chain logged, LightNet v2 packet interlock. Rust refimpl running.
Rust · OSS · 35 KB
NovaHoney
A honeypot that classifies intent, not just signatures.
Concept-grid firmware. Classifies attacker intent, writes payload-free logs safe to deduplicate. No CVE database required.
Custom builds · consulting · from $499
Need an AI tool
built for your workflow?
Bring me an idea, a messy process, a dataset, or a repeated decision. I’ll bring the same thinking that produced Lumina OS, NovaGlyph, and STONE to your specific problem — from a quick audit to a finished product.
Workflow Audit
I map what should be automated, what should not, and what the first tool would look like.
Prototype Build
A working AI prototype for one specific workflow — local or cloud.
Product Build
Packaged app, offline tool, internal dashboard, or digital product with checkout and download.
Custom Specialist
Domain-specific expert system binary. From $499. You own the source.