Independent Development Studio
We build AI platforms, industrial field tools, mobile apps and games — from a food-rescue network serving Alberta to a multi-provider AI desktop suite that runs entirely on your own machine. Eleven products. Six platforms. Every line shipped.
Measured, not claimed
Every figure below is counted directly from the source tree — excluding dependencies,
build output and generated files. Re-run npm run stats to regenerate it.
The portfolio
Everything we've built — tagged by where it actually stands. Nothing here is a mockup or a landing page waiting on a product. Production means live and in users' hands. Beta means installable and working, still being hardened. In development means it runs, with work remaining. R&D means we're proving the idea out.
A food- and supply-rescue network that connects surplus to the people who need it, across Alberta and Canada. Agentic matching pairs donations with shelters, food banks and crisis services in real time. Ships as a web app, an installable PWA, and native iOS and Android builds off one codebase — with its API running on Cloudflare Workers at the edge.
Five AI providers in one Windows app — Gemini, Claude, OpenAI, and local Ollama or LM Studio servers, each with live latency on a provider pulse. It calls those providers directly: no server of ours in between, no account, nothing synced. Keys are sealed with Windows DPAPI and the interface only ever sees the last four characters. Grounding tools, a local IDE with a coding agent, an executing workflow graph, audit logs and snapshots that survive restarts.
Flat-pattern layout for industrial insulators — gore elbows, tees and laterals developed to the ordinate, with material take-off, ASTM reference tables and an AI field assistant. Built for Alberta SAGD steam distribution work and Red Seal compliance.
Boondocking spots, ghost towns and power hubs across North America — plus Sky Scan, which turns the phone camera into a measured satellite obstruction map by solving lens bearing and elevation from the rotation matrix.
A full DAW built around generative composition — a Lyria-driven track builder, multi-track arranger with an AI section editor, six-knob macro rack, mixer console with per-channel sidechain and VST slots, spectral analysis and a master bus. Exports audio, video and MIDI.
A vertical climb through generated dimensions, with synchronised multiplayer sectors that show other pilots' altitude live on your HUD. Describe a universe in plain language and the rift engine builds it — heavy gravity, neon platforms, whatever you ask for. There's an in-game Oracle you can genuinely consult on run strategy, a prestige ladder that trades progress for permanent physics multipliers, and a tiered emporium of trails and avatars.
Turns the coding agent you're already talking to into an orchestrator. One tool call puts a task to five specialists at once — architect, optimizer, sentinel, builder, executor — each backed by its own local Ollama model. They aren't one-shot prompts: every member gets read, list and search tools and investigates the workspace itself, so reports cite real code. They run concurrently, so consulting five costs about what consulting one does, and only the orchestrator may write — a member reaching for a write tool is refused, which is why the whole team also works in plan mode.
A genre mashup prototype — first-person shooting layered over cooperative tower defense, with networked multiplayer and a neon-horror art direction.
Real screens, real builds
No mockups — the screens below are captures from the running applications on a working machine, alongside the film we cut for the field.
Set genre, mood, key, vocal style and groove, then let a Lyria model build the track — and take over the moment you want to. Everything after generation is a real DAW.
Gemini, Claude and OpenAI when you want them — Ollama and LM Studio when you don't. Requests go straight from your machine to whichever you pick, with no server of ours in between, no account and nothing synced. Run it entirely on local models and nothing leaves the box.
Insulation work happens on pipe racks in the cold, in gloves, one-handed. That's the constraint the whole product is designed around — and the reason the maths has to be right the first time.
The speed control
Most AI tools give you one gear, and it's the slow one. Gemini's thinking models deliberate before they answer — brilliant on a hard problem, agonising when you already know what you want. QuantaSync puts that on a slider. Drag it to the last stop and the model stops deliberating entirely: it answers the instant it can.
This panel is a recreation of the app's behaviour, timed from measured runs — the page makes
no model calls of its own. Inside QuantaSync the slider sets Gemini's
thinkingLevel, and the
last stop sends MINIMAL
rather than omitting the config — on a thinking model, leaving it out means "model default",
which is not the same as asking it to stop.
Live in your browser
Three of our products, running right here. The calculator below performs the same lobster-back development an insulator does on the shop floor — real trigonometry, real numbers, no video.
| Piece | Segment angle | Throat | Heel | Centreline |
|---|
Standard lobster-back development for a 90° elbow: end pieces subtend half the angle of the intermediate gores. Throat and heel are struck from the elbow centreline radius offset by half the insulated outside diameter — throat = (R − OD/2) · tan(β/2), heel = (R + OD/2) · tan(β/2). Add the lap once per circumferential seam.
A visualisation of how the orchestrator actually fans out: one task goes to five specialists at once — Mythos on architecture, Gemma on the shortest correct path, Llama hunting races and leaks, Mistral writing the code, Granite ordering the steps that verify it. Each runs on its own local Ollama model and reads the workspace before answering. Because they run concurrently, total latency is the slowest specialist rather than the sum of five.
This is real linear algebra, not an animation. Eight complex amplitudes are held in a state vector and every gate is applied as a unitary on that vector — Hadamard, Pauli X/Y/Z, phase S and T, and a controlled-NOT that entangles two wires. Bar height is |ψ|², the probability of measuring that outcome; bar colour is the amplitude's phase, which no measurement can see but which decides how the paths interfere. Entanglement is computed properly, from the purity of qubit 0's reduced density matrix after tracing out the other two: 1.000 means separable, 0.500 means maximally entangled.
State-of-the-Art Architecture
Five browser capabilities, each running against a real Web API rather than a mock: WebGPU tensor compute, a language model resident on your own GPU, a 3D spatial visualiser, measured edge latency, and the Web Speech API. Open DevTools while you use them — that's rather the point.
A real WGSL compute shader dispatched to your graphics adapter, running the single-precision matrix multiply that every neural network layer is built from. The same multiply is then recomputed on the CPU and the two results are compared element by element — so the speedup on screen is measured on your hardware, not quoted from ours.
// Pick a matrix size and press Run. The shader below is dispatched to your GPU,
// then the same multiply is recomputed on the CPU and the two are compared
// element by element. Nothing here is pre-written output.The card above runs the primitive. This runs the whole stack: a real transformer, loaded into your GPU and executed here. Weights download once from Hugging Face and are cached by the browser — after that it works with the network off. Open the Network tab while you chat and you will see no requests, because there is no server to make them to.
Nothing downloads until you press Load. The model is a few hundred megabytes — small enough to be quick, small enough that it will sometimes be wrong. That is the honest trade at this size, and it is running on your GPU either way.
Interactive 3D orbital canvas representing QuantaCore platform nodes. Click nodes to inspect runtime specs or drag to orbit in 3D space.
Live edge endpoint latency, browser heap monitoring, and hardware thread concurrency stats.
Control website navigation and hear studio status updates using real-time voice commands.
"Go to work" "Run local ai" "Telemetry" "Contact"
What we do
One studio taking products from first commit to signed installer — across desktop, mobile, web and the edge.
Multi-provider gateways, local model meshes, streaming agents and orchestration that runs on your own hardware — not just an API key wrapped in a chat box.
Kotlin and Compose for Android, Capacitor for cross-platform delivery, and the hard parts done properly — CameraX geometry, sensor fusion, offline-first Room caches.
Electron applications packaged into signed NSIS installers and portable builds, with OS-level credential vaults, real telemetry and IPC that respects the main thread.
Cloudflare Workers, Pages and D1 at the edge; Firebase and Google Cloud where it fits. Secrets stay server-side, keys never reach the bundle.
Interfaces that hold up in a boiler room and on a phone at dusk. Dark-first, high-contrast, one-handed — designed around the conditions the work actually happens in.
WebGL and GPU shaders on the web, networked multiplayer prototypes, and real-time simulation — including the raymarched shader running behind this page.
The studio
QuantaCore Labs is an independent development studio. Small, fast, and unusually unwilling to ship something that doesn't work.
Every product on this page exists as running software with an application ID, an installer or a store package. We label R&D as R&D. Nothing is announced before it runs.
QuantaSync keeps keys, conversations and logs on your own disk and talks to providers directly — no account, no server of ours in the middle. Sovereignty Suite blocks trackers without reading your history. When something can run locally, it runs locally.
Camera field-of-view geometry, lobster-back development, non-blocking multi-model orchestration. We take the problems where getting it approximately right isn't good enough.
A food-rescue network for shelters. A layout tool for insulators on a steam line. Software aimed at real work, not at a demo reel.
Get in touch
Whether it's a mobile app, a desktop platform, an AI system that has to run on your own hardware, or a field tool nobody has built yet — tell us what you're trying to make.