We're an independent AI R&D lab. We design and ship our own products, release open-source tools, and fund our own research — spanning a from-scratch operating system in Rust, real-time computer vision, generative media, and self-improving reasoning engines. We also build production software for clients.
We show you everything. Even the parts that make us look stupid.
Your data is yours. We don't sell it, mine it, or "anonymize" it into a product.
Research, tools, and libraries — out in the open. Some products are commercial. We're transparent about which is which.
No enterprise gotchas. No surprise invoices. Transparent pricing or bust.
We hack our own stuff before you do. And if you find something first, we pay you.
Built by humans and their opinionated AI. Contributions welcome, egos optional.
An infinite, self-playing cyber range.
A procedurally-generated, verifiable cyber range paired with a tiered red/blue agent that learns to attack and defend through self-play — the range is the training-data factory, the RL reward signal, and the benchmark all at once. It also ships a browser terminal where you practise recon → exploit → pivot → loot against live-flavoured scenarios in a safe simulation.
Enter the RangeIntent in. Running software out.
A universal intent runtime and software fabric. Instead of writing line-by-line code, you declare goals, constraints, and contracts — and a multi-objective planner compiles them into a content-addressed execution graph that runs across targets. Built in Rust around seven typed primitives, with native telemetry, capability-based security, and a continuous optimisation loop.
View on GitHubAI roleplay & living worlds.
Create characters with depth, memory, and real opinions. Build immersive narratives, explore persistent living worlds, and generate stories with AI-powered companions. Think Character.ai — but bigger, better, and with actual creator tools.
Visit ChazzAIReal-time GPU vision: plates & faces.
A modular, GPU-accelerated computer-vision stack. Reads licence plates from any angle through extreme degradation, tracks them with ghost-resistant persistent IDs, and runs reconstructable FaceID with age, gender, and emotion — all locally via onnxruntime-CUDA.
Brief in. Platform-ready ad out.
AI-powered ad production pipeline. Feed it a brief, get back a full video or static ad — storyboard, visuals, voiceover, background music, and final assembly, automated end-to-end. Built-in compliance engine rewrites edge-case briefs into tasteful creative that passes Meta, Google, TikTok, and Reddit automated review.
Private, in-browser file tools. Zero uploads.
Convert, compress, and modify files — plus encoders, hashers, and generators — running entirely on your device. An any→any file engine wrapped in a growing library of instant utilities. Nothing is uploaded; your data never leaves your browser.
Launch WebToolsSports league voting, done right.
Multi-tenant 3-2-1 player voting for sports leagues. Best-on-ground picks, live results, and season leaderboards — available on web, iOS, and Android from a single codebase. No credentials required for voters, full admin controls for league owners.
Visit MatchPointsAI testing with persona agents.
Unleash a swarm of AI agents with distinct personas onto your web app. They browse, click, judge, and report back — like a QA team that never sleeps and has very strong opinions about your button placement.
Computer vision fire & smoke detection.
Real-time 3-tier detection pipeline using YOLO, CLIP, and depth estimation. Context-aware hazard scoring, spread prediction, and annotated live overlays. Built for environments where false positives are expensive — and missed detections are catastrophic.
Raise money the simple way.
A 100-square board where each square is a chance to win. Simple rules, transparent odds, zero shady fine print. Perfect for fundraisers, communities, and anyone who believes raising money shouldn't require a PhD in fintech.
Visit Lucky 100These are sites we designed, built, and shipped for other people — clients who trusted us with the real thing. We don't own them; they're work, not our products. For the things we build and run ourselves, see our own projects up the top.
From peak sporting bodies to charities — organisations that trusted us to build the real thing.
Cognitive Industries is a tight crew of engineers and our opinionated AI, building across an unreasonable surface area — from a from-scratch operating system in Rust, to heterogeneous compute that targets GPUs and quantum backends, to self-improving reasoning engines, real-time vision, and the products you've already scrolled past. We ship research-grade systems and production apps with the same standard: do the hard thing properly, or don't do it.
A representative sample — not an exhaustive list, and nowhere near a ceiling. Across hundreds of projects we’ve shipped with far more than what’s below, and we pick up whatever a problem demands. If it compiles, runs, trains, renders, or breaks — we’ve probably had our hands in it.
+ 481 listed above — and a long tail beyond it. The stack is never the limit.
A sample of the research and systems work behind the scenes. Not all of it ships as a product — some of it just had to exist.
An experimental operating system written from scratch in Rust — capability-based security, deterministic replay, and a content-addressed object graph as first-class primitives. Runs on-device ML at ~60k inferences/sec.
Rust · bare-metal x86-64"Write once, run on any accelerator — including QPUs." A provider-agnostic IR that plans whether each piece runs on CPU, GPU, or quantum hardware, with cost estimation across IBM Quantum, AWS Braket and D-Wave. 204 tests, 91% coverage.
Python · Qiskit · Braket · OceanA self-improving engine where a symbolic NARS core reasons autonomously and only escalates to LLMs when stuck — then compiles those rescues into permanent knowledge. Proved 6/7 theorems and rediscovered 3 physics laws, 50% LLM-free.
Python · PyNARS · SymPy · PySRText-to-video on an 8GB GPU where frames never touch disk as plaintext — a cascade-encrypted, deniable in-memory vault with anti-capture viewing, optionally bursting to confidential H200 compute.
Python · Diffusers · XChaCha20A 9.6M-parameter model trained from scratch that beats far larger models once paired with test-time search and verification — a 7× lift at scale, with the negative results documented as carefully as the wins.
Python · PyTorchTests whether intelligence emerges from many small specialised models behind a learned router — Mixture-of-Experts meets Global Workspace Theory, with routers that learn domain from loss alone and variable-length reasoning paths.
Python · PyTorchEvolves all seven components of an LLM stack at once — preprocessing, tokeniser, architecture, training, inference, post-processing, representation — with a UCB1 bandit choosing what to improve so gains compound on a shared genome.
Python · OpenEvolveFounder · Lead Engineer
His introduction to programming was a glowing MS-DOS prompt as a kid, handed to him by his uncle Dominic White — a Sydney-based engineering leader with two decades of major platform transformations across Australian broadcast and streaming (a year-long SBS replatform with a team of 50, engineering teams of 50–60 built at SBS and Switch Media) and founder of alt.script. The spark caught and never went out. He built his first robot at twelve.
He dropped out of high school — and a month later was hired back by the same school to run its IT. Everything since has been the same pattern: bet on the hard, interesting thing. Even sleep doesn't switch him off — he dreams in code and writes it up the moment he wakes; away from the keyboard, it's still the only thing on his mind. Nothing boring holds his attention; he'd rather crash and burn on something ambitious — and has, many times — than commit to a comfortable, boring life or a comfortable, boring codebase.
He's entirely self-taught. The one time he tried formal training — a TAFE certificate here in Australia — he finished the whole course, final project included, inside the first three days of enrolling, without AI. They wouldn't accept it all at once: staying meant months more attendance, stacks of non-code paperwork, and in-person oral exams. He had no patience for any of it, so he left and went back to shipping.
He works until he physically can't, then recovers and does it again. The burnout is real; so is the supplement stack he engineered to climb out of it in days rather than weeks.
We're not a fully open-source shop — some products are commercial, and that's how we keep the lights on. But we open source what we can, share our research freely, and welcome contributions. Closed source never means closed community.
Universal intent runtime & software fabric. Declare goals, ship a content-addressed execution graph.
GPU-accelerated real-time vision: any-angle plate reading, ghost-resistant tracking, reconstructable FaceID.
AI-powered ad production pipeline. Brief → storyboard → video, end-to-end.
3-tier computer vision pipeline for real-time fire and smoke detection with hazard scoring.
Multi-tenant sports league 3-2-1 voting system. Web + iOS + Android.
AI agent swarm platform for automated web testing with persona-driven exploration.
Every cent goes straight into improving what's already here and shipping more & better open source — more compute, more research, more released code. None of it reaches the developers. We build for the love of it; donations just buy the open-source work more runway.
We're small. We're weird. If that's your thing, say hi.
We believe the best teams are built on curiosity, not credentials. If you're the kind of person who reverse-engineers things for fun, we should probably talk.
▚ // recruit clearance terminal detected — initiate()
We owe these legends money. This is our public shame.