Sydney Quant Pty Ltd · Australia
Sydney Quant is a proprietary trading firm based in Australia. We design, test and operate automated market making systems with our own capital. Alongside the desk is a small software studio that builds apps, games and data tools, and six of those tools run live on this page.
Division 01 · Markets
Sydney Quant is a proprietary trading firm. We trade our own capital with automated market making systems, and we build and run every part of the stack behind them, from data capture through to risk.
The core of the business is market making: quoting both sides of a market continuously, holding inventory for short periods, and earning a small edge over a very large number of trades. Turnover is high by design. Because the edge on any one trade is small, the whole business rests on being right on average, and on knowing that with confidence rather than assuming it.
That shapes how research is done here. Every new feature, strategy or configuration change is run as an experiment, with a hypothesis, a defined test and a bar it has to clear before it goes anywhere near live capital. We look for plateaus rather than peaks: a parameter that only works at one setting is treated as noise. A signal has to be statistically significant, has to stay significant out of sample and after costs, and has to be clearly distinguishable from luck before we call it a signal. Systems that pass are then tested for robustness, because something that only works in one regime is not a system.
A large share of the day-to-day work is data. Clean, consistent and current datasets are the foundation everything else sits on, so we capture our own, keep the raw record untouched, and maintain the cleaned versions with the same care as the trading code. A good result on a bad dataset is worse than no result.
The business is scaling up in 2026, with more capital at work, broader coverage and infrastructure built to carry it. Everything described on this page was designed so that growth means turning things up rather than rebuilding them.
Four tests every change has to pass before it trades. The same four, every time, whether the change is a new strategy or a one-line tweak to a config file.
Every parameter is swept across its range, and we keep the configurations that sit on a broad, stable plateau. A setting that only works at a single point is discarded, however good that point looks in the backtest.
A signal has to be statistically significant on a sample large enough to mean something, and the test has to account for how many ideas were tried before this one. Profitable in a backtest is not the same thing as evidence.
Data is split into fitting and holdout sets, and nothing is re-tuned to the holdout. Fees, spread, slippage and latency are modelled rather than assumed away. An edge that does not survive that was never an edge.
A system has to hold up across market regimes and across reasonable perturbations of its inputs before it trades real money. The question asked of every result is the same one: can we show this is not luck?
Six stages, one system. What the models look for stays private. How the system is engineered and operated is the part we can describe, and it is the part that decides whether a strategy survives a live venue.
Data is captured continuously and stored raw before any cleaning, so a decision made last month can be checked against exactly what the system saw at the time. The cleaned datasets are built from that record, refreshed on a schedule, and treated as a product in their own right.
Research and production run the same code. A backtest is the live engine fed historical data through the same interface, so a result in research is the result you get in production. Every experiment is recorded with its hypothesis, its data range and its outcome, whether or not it passed.
Models are re-evaluated on a schedule against current data, and every output is stamped with the inputs and the version that produced it. A signal that cannot be traced back to its inputs, or that has drifted from the behaviour it was tested on, is not allowed to size a position.
Orders are sized from the model’s confidence and the account’s risk budget, then routed with idempotency keys so a retry after a timeout cannot become a second position. At high turnover the unusual cases are routine, so rejects and partial fills are handled as normal paths rather than exceptions.
Limits on position, exposure, order rate and drawdown are checked before submission. Positions reconcile against the venue on every start, every reconnect and on a schedule in between. A mismatch halts trading until a person has looked at it.
Every session writes a log detailed enough to reconstruct it tick for tick. When something odd happens it is replayed rather than debated. Alerting covers gaps, rejects, drift, and silence where there should be traffic.
Division 01 · Markets
Working versions of tools we use internally, cut down to fit a browser tab. Everything computes on your device and no requests leave the page.
Monte Carlo · parameters
Footylab · edge rules
Break-even probability at these odds is —. Below that, the price is against you no matter how good the pick feels.
Portfolio Lab · funds
Pick an equal-weighted portfolio. Real parsed holdings from both issuers.
Disclosure coverage of each fund’s top 25 lines: —
Alongside the trading desk, Sydney Quant builds consumer software: apps on the App Store, games, websites, data tools and generative art. Everything below was designed, built, released and supported in-house, to the same standard as the trading stack.
Division 02 · Flagship
Auto Arena is the studio’s flagship game: a ranked auto-battler built on a deterministic engine, with an authoritative match server, a balance harness that tests every unit against every other, and a video pipeline that cuts trailers from real matches. It is free, runs in a browser tab and needs no download.
Fifteen seconds of a real ranked match: plan, star up, horn, 3–1. Every frame is captured from the running game by a rig that replays the match on a fake clock, so what you see is the actual build.
Draft a hero and four minis. Spend an elixir budget placing them on your half of a 7×8 board while the enemy’s half sits under fog. Last round’s army shows through as silhouettes, and whatever they add this round does not. Drop a duplicate on a mini and it merges up a level. Then the horn goes and nobody touches anything.
The battle resolves itself: nearest-enemy targeting, abilities charging off damage dealt and taken, a mini that never got to cast firing its spell with its dying breath. First to three rounds wins, armies persist and grow the whole way, and a round lost is ground that does not come back on its own. About three minutes, start to verdict.
Trophies, eight leagues with floors you keep for the season, monthly resets and an account level that never does. Pick a name and you are playing: no email, no password, nothing to install. Attach a login later and the same account follows you to your phone.
Thirty seconds to plan it. Both sides plan at once, and neither sees the other.
Eight minis and three heroes to choose from, four slots and a cap of two copies, so every army is a mixed comp. The hero costs nothing, never stars, and can be walked anywhere on your half every single round.
Both players draw the same 36 elixir across a match, paid out a round at a time and tilted toward whoever is behind. Unspent elixir carries. Eight bodies is the ceiling, so after that the only way to spend is up.
Drop a duplicate on a mini and it merges: ×1.65, ×2.3, and a ★★★ at four times its starting stats for three times its cost in one step. It is deliberately inefficient, and deliberately the biggest body on the board.
The horn goes, the fog lifts on what they added, and a fixed 20 Hz simulation plays it out. There is no scrubbing and no micro. The plan was the whole game, and about half a minute later there is a verdict.
Every unit is drawn in code, rigged and hand-animated. These are rendered by the game’s own renderer, run here at portrait size. Each one charges mana by hitting and by being hit, and casts the moment it fills.
Frenzy. +80% attack speed and +30% damage for four seconds. The cheapest body in the game.
Piercing Shot. 250% down a corridor. Fires faster the longer the round runs, which is what stops a grind drifting to the timer.
Shield Wall. absorbs 35% of its own health for six seconds. Cast with its dying breath, it can survive the blow that killed it.
Skull Slam. 200% all around it plus a stun, and a stun lands before its damage, so it cancels the dying breath it would have triggered.
Charge. barrels up its own column the instant the horn goes: 300%, a tile of knockback and a stun. Read the lane wrong and it hits the wall and stuns itself.
Shadow Leap. vaults the whole field at the horn and lands behind their line, untargetable in the air. One thing on the board can touch it: a Juggernaut charging up that column.
Arcane Blast. 220% in a radius. The only unit with a real price on it: about ten wins’ worth of coins.
Radiance. one 400% heal across every wounded ally. All of its healing is in that cast, so keeping it alive until it casts is the whole skill, and killing it first is the whole answer.
Groundbreaker. 250% and a one-second stun around it. Heroes deploy free and grow 15% a round, so by round five they are a different unit.
Arrowstorm. 180% to every enemy within five tiles at once. Put her behind a wall and she decides crowded rounds.
Aegis. shields itself and every ally within three tiles for a quarter of their health. A whole front line that has to be killed twice.
The parts a player never sees, and the reason the parts they do see can be trusted.
The engine is headless and touches no DOM, no timers and no wall clock. The same armies and the same seed produce a bit-identical battle on any machine, which is what makes a replay exact rather than a re-enactment.
Both sides plan at the same time and neither sees the other. The server validates every army, runs the battle itself and sends out the deployments and the seed; each client replays it and checks its own verdict against the server’s.
A harness plays every legal loadout against every other on shared seeds at three skill tiers, and a unit is only balanced when swapping it into the same deck moves the result by under three points. Fairness is held to the same standard: identical armies rotated 180° read 49.8 / 50.0 over 60,000 battles.
Each video here starts as a spec file. One command plays a real match through the real interface on a fake clock, captures only the frames the edit needs, cuts them on the music’s beat grid and masters the result. Change how the game looks and the footage is captured again automatically.
Trophies move on an Elo-style curve: +30 / −20 between equals, up to +48 for beating someone far above you. Eight leagues, and reaching one locks its floor in for the season, so a bad evening cannot undo a good month.
Type a name and the next screen is a match. Matchmaking opens a trophy window and widens it until it finds someone; ten seconds with nobody there and you get a bot near your rating instead of a queue.
Division 02 · On the App Store
Sum Hero was the first game to go through App Store review. Model, backend, client, store listing and support were all built in-house. It is free on the App Store and playable in the browser.
Two players get the same questions at the same time, and it is settled on speed and accuracy. Points come from answering fast and from stringing correct answers together, so a clean run beats a careful one. That is what makes the last ten seconds worth staying for.
Ratings move on results, matchmaking pairs you against players near your level, and the leaderboard resets weekly so a good week actually counts. Play as a guest with no account at all, or sign in so a rating follows you between devices.
No ads, no advertising identifiers, no data sales. The app collects what matchmaking and rankings need and nothing beyond it. The privacy policy lists every field, and deleting your account takes one email.
Division 02 · The lab
The audio visualiser is a browser port of an offline renderer we built for music videos, with the same layer model and colormap. The tax tool runs the same bracket schedule as the shipped app. The fractal explorer is here because it was fun to write.
Audio visualiser
Kicks detected: 0 · tap the canvas to add one
Audio is decoded and analysed in your browser. Nothing is uploaded, stored or sent
anywhere.
Aussie Tax Calculator
| Bracket | Rate | Income in band | Tax |
|---|
Fractal explorer
The Mandelbrot set and three of its Julia relatives, computed pixel by pixel in your browser. Click anywhere to dive in; hold Shift and click to back out. Infinite coastline, four kilobytes of JavaScript.
The catalogue
Shipped products, private infrastructure, and apps that were released and later taken down. Status is checked against the App Store rather than claimed. Open any card for the full write-up, or press ⌘K to search.
How we work
These are the review criteria applied to every change, in the trading stack and in the apps alike.
Code is read far more than it is written, and debugged at the worst possible moment. Boring, obvious implementations beat elegant ones that need a mental run-up.
Any number a system produces should be traceable to the inputs that made it. If a result cannot be explained in plain language, it does not ship.
Anything touching money stops rather than guesses. Stale data, a lost connection or a breached limit halts the system; nothing resumes without reconciliation.
Model, backend, client and store listing. Owning every layer means the second product costs a fraction of the first, and no vendor gets to decide the roadmap.
The company
Sydney Quant Pty Ltd is an Australian private company. Its main business is proprietary trading. Its second is a small software studio that designs, builds, ships and supports the products on this site. Research, engineering, design, release management and support all happen in the same place.
The work is published under the name cheesepickle, which is also the name to use if you write to us. There are no headshots or bios here. The catalogue above is the record of what has been built, and the live tools are there so you can judge the work for yourself.
Not everything survives. Three apps in the catalogue were built, released and later taken down, and they are labelled that way rather than left looking live. The two that are still on the App Store are linked directly.
We use AI tools to speed up parts of the build and research process, in the same way we use any other tool. The products, the research and the decisions behind them are our own.
Trading work is discussed at the level of systems, infrastructure and research method. Strategies, signals, instruments and parameters are not published.
Contact
Support requests, bug reports, data or infrastructure work, or a question about something in the catalogue. One inbox, read by the person who built the thing you are asking about.
