Aqua · the engine beneath Zeton
When hardware stops getting faster, software has to get smarter.
Aqua is the engine underneath Zeton. It reads the machine it runs on and adapts to it, so a firm can serve far more client portfolios on the hardware it already has.
What it means for a firm
Personalization is a computing problem. Aqua makes it a cheap one.
More portfolios on the same machines.
Aqua turns the hardware a firm already pays for into far more client portfolios served, which is what makes personalization affordable.
The same answer, every time.
Aqua changes how fast an answer arrives, never the answer. Results match the reference to the last digit a computer stores.
Nothing to configure.
No tuning, no special hardware, no flags. Aqua discovers the machine when it starts and adapts on its own.
One engine, many machines
The same code, a different route on every machine.
Aqua profiles the hardware when it starts, classifies what it is working with, and routes each operation to the strategy that machine can run fastest. Pick a machine and watch the route change. The caller never writes a line of configuration.
Aqua profiles the machine
Aqua detects constrained resources and selects a conservative parallel strategy.
and routes the work
On a constrained cloud, Aqua routes to conservative parallelism and cache-blocked kernels.
Nothing is configured by the caller. Your code never changes; Aqua re-routes underneath.
Three things have to be true at once.
It decides at runtime.
The machine is discovered when the program starts, not assumed when the code was written.
Aqua also carries a declarative layer: you state what to optimize and the constraints it must respect, rather than how to compute it. That is the surface our patent application covers.
Measured, not modelled
Correctness first. Then speed.
No timing counts until the answer is proven identical. Every figure below passed that gate first: results agree with the reference to within 2.4 × 10⁻¹⁶.
A platform workload
Many clients, one date
- 56.4M
- client accounts optimized per machine per hour, up from 9.75M
- 5.78×
- more client accounts served per machine
- 17.3 → 6.6 ms
- for the slowest single account
1,000 client accounts at full machine load, on three cloud machine types, against the same workload without Aqua.
The engine, one operation
In isolation
- 10.32×
- CPU efficiency on a warm solver pool
- 9.49–12.29×
- held across a 15× range of machine sizes
- 1.3–1.9×
- on general numerical workloads, across 209 tests
Against our own pre-Aqua implementation, at a small universe size. It shows the engineering is stable; it is one operation, not a firm’s whole workload.
Our production backtest
One strategy, ten years
- 5–15×
- faster than the implementation it replaced
- 36 of 36
- configurations with every financial metric matching
Against our own pre-Aqua implementation, not a third-party system. Not all of the gain is Aqua’s alone, and the upper end depends on available cores.
Three different workloads, measured separately against different baselines. The figures are not comparable with each other, and they measure computation, not investment returns.
Protected
Born in NYU research. Licensed to Aquon alone.
- Exclusive licence
- Exclusive, worldwide and in all fields, with the right to sublicense. Executed with New York University in August 2026.
- Patent application filed
- A US provisional application filed through NYU in June 2025, followed by an international (PCT) application in 2026.
- Invented here
- Our CTO, Carlos de Oliveira, is a named inventor on the application.
See what Aqua does for your book.
Talk to us about running Zeton, or about licensing Aqua inside your own platform.