CPU, not cloud
Every architectural decision in Qovaryx serves one constraint: it has to run on your processor, on your machine, locally. Not “could in principle”. Not “runs if you also pay for our cloud”. Runs.
Ch. 01 — Engineering
Why this is the right constraint
- Latency. A network round-trip is an eternity when option pricing is moving. Local evaluation returns without leaving the machine.
- Cost. If every evaluation cost us metered API calls, the price of the product would have to reflect it. Running locally is free at the margin, for us and for you.
- Privacy. Your charts, your positions and your decisions never touch our infrastructure, because there is no path by which they could.
- Reliability. Our infrastructure being down does not stop you working. There is nothing of ours in the loop to be down.
What it forced us to give up
The honest answer is depth. A very large general model would have far more raw capacity than anything that fits on a desktop processor. We do not have that, and no amount of engineering conjures it.
What we have instead is a cluster of small specialists that, between them, cover the dimensions that actually matter for this task. The total size is a fraction of a frontier general-purpose model, but the surface relevant to the job is dense rather than thin. Narrowness is what buys back the capacity.
Cloud inference is what you do when the model is too big for the machine. We made the machine sufficient by making the model small.
On the direction of the tradeHow we made it fit
We are not going to write out the recipe — that part took a year and a half. The shape of it:
- Gradient-boosted classifiers for the chart specialists. Compact enough to sit inside the application, with no graphics card required.
- Small neural specialists for the contextual work — market character, scheduled events, the dedicated objector. Optional, opt-in, and processor-only by default.
- Feature preparation written in plain Python. No heavyweight deep-learning framework sitting on the path that has to be fast.
- Cached broad-market context with a short expiry, so repeated evaluations do not repeat the same fetch.
- Specialists loaded lazily — each pages in the first time it is actually needed rather than all at once at startup.
What it looks like while running
The engine panel in the application reports processor and memory use as it works. There is no graphics-card row, because there is no graphics-card dependency. There is no spin-up delay while remote capacity is provisioned, because nothing is being provisioned. The first evaluation returns in the time it takes the machine in front of you to do the arithmetic.
We publish no performance figures. This page is about where computation happens and why. Any return, win-rate or accuracy number attributed to Qovaryx did not originate from us.