Systematic, AI-augmented trading across global markets.

02 · Executive summary

The industrialised loop around strategies.

Seven years building one platform that researches, validates, deploys, executes and continuously rebuilds portfolios of systematic strategies — across every major asset class, in every market above.

What makes it different is not any single strategy. It is everything around the strategies.

Five books run today: thirteen strategies live over Hong Kong, US, Japan and Singapore index futures, executed through two brokers; twelve in paper on China index, commodity and energy futures; and three new lines this autumn, all on paper — eight currency futures traded by their own desk and, since 17 October, by the first two systematic legs; eleven global commodity and Treasury micro futures traded by five class desks while their first campaign runs; and a China A-share desk on CSI 300 whose larger orders go through the algorithm engine, beside three basket policies still in research. Inside the books, nineteen specialist agent desks manage positions under hard deterministic limits.

We are seeking partners who want exposure to a repeatable process rather than a single strategy: the same pipeline that continuously re-validates and redeploys the live portfolio can absorb new markets, new capital and new signal sources without changing shape.

04 · Platform at a glance

One loop, four surfaces, one core.

Four surfaces, one core. Everything beneath the surfaces executes, risk-manages and reports the same way for every market and asset class.

Deployment topology

Two clusters, three boxes each, one terminal in front of both.

Design principles

• Config-driven, no hard-coded instrument logic. Adding a market is a config file, a contract specification and a feed connector — not a code fork.
• One code base for backtest, paper and live. The same strategy class runs in all three modes; only the header config changes.
• Deterministic core, swappable intelligence. Execution, risk and journaling are deterministic and unit-tested; models sit behind narrow contracts and can be replaced or switched off without touching the pipeline.
• Everything observable. Heartbeats, pattern-state exports, NAV logs, alerts, daily reports, and a browser terminal on desktop and phone.

06 · The research factory

A pipeline designed to say no.

Candidates enter on the left. Each gate is independent of the others, and a candidate is rejected by the first one it fails.

The most recent candidate matrix

147 cells, partitioned by outcome. A cell is rejected by the first gate it fails.

Portfolio constructor

Seven weighting methods, a correlation cap between legs, and a search for out-of-sample performance subject to a full-sample floor. The same pass sets each candidate’s daily loss clamp.

Data plumbing that keeps research honest

• Two cache sets per market, refreshed weekly — one for testing, one for training.
• Sources are never mixed. Every dataset carries its origin.
• Hourly re-runs keep every candidate comparable to the live book.
• A splice that is reported, not hidden. The FX research history is a decade of spot bars joined to the futures’ own bars; volume-driven signals train on the futures era only, and every FX candidate shows how it behaved on each side of the join.

08 · AI layer

From macro filter to autonomous trader.

8.1  The investment committee — the book argues with itself

A macro filter, not a signal generator. Before each market open the books trade — nine sittings every weekday — an adversarial committee argues over every market, and the verdict becomes a small machine-readable bias the strategies read on every bar.

• The cast. Regional analysts market by market; specialists on macro, rates, FX, commodities, agriculture, liquidity, sentiment, event risk and geopolitics; advocates over two rounds, a red team paid to attack the consensus, a risk agent, a cross-asset checker and a synthesiser — each able to run on a different model.
• What they see. 152 engineered indicators, the calendar with its event-risk windows, the regime state, the news flow in English and Chinese, and since October an FX tier of positioning, fixings, yields, policy rates and event feeds.
• What they produce. Per market: weights by signal type, a vote with conviction, and the agreement a signal must clear before it may act.
• A view on every currency. Each sitting now ends with a dedicated FX step that writes a bull/bear view, with conviction, on each of eight currencies against the dollar. A view is stored on the currency and flipped at read time where a contract is quoted the other way round; if the FX data is missing the step is skipped rather than guessed. A ninth sitting, an hour before the London open, was added for it.
• A view on every commodity class, at home and abroad. The main verdict also writes a stance on five global commodity classes — precious, base metals, energy, agriculture and rates — beside the seven China commodity groups, so a Shanghai gold desk and a COMEX gold desk read the same committee from their own side of the market. Each sitting now writes thirty-two views; a class left out is seeded from its reference contract’s day move, never invented.
• Fail-safe by construction. No bias file means no change; a stale file is dropped rather than trusted, so a silent upstream failure degrades to plain systematic behaviour. The layer can be switched off per instrument without touching a strategy. It was armed on every live leg in September 2026, and the same file is read by the agent desks.
• It has to earn its keep. A cycle is skipped rather than half-applied when calls fail, and every cycle is scored against what the market actually did.

8.2  Agent Trader — a model that runs a book, safely

The model writes a decision. Everything that turns it into an order is ordinary, tested code.

• A stop on every open, enforced by both the executor and the monitor, inside per-instrument, cluster and margin caps it cannot argue past.
• Paper to live is a file edit by a person. No button, API call or model output can promote an agent.
• It is told in advance what it may not do — and how often it proposes it anyway is measured and shown beside P&L.
• Memory is an auditable note store, not a growing prompt: the model proposes, a nightly pass writes lessons under evidence floors, operators pin instructions.
• Every Saturday it reviews its portfolio’s week, not just its own trades.
• Bound to a systematic book, each management action sits on its own shadow, approve or automatic rung, promoted one at a time by an operator and reversible the same way. Every operation began in shadow, journalling what it would have done without acting; as the record accumulated each was promoted, and as of September 2026 all of them run automatically. What bounds a desk is not an approval queue but the mandatory stop, the per-instrument, cluster and portfolio-wide caps, the margin ceiling and the circuit breaker.
• New rules ship switched off, evaluating and journaling what they would have done until the measurement justifies arming them.
• Profits on managed legs are protected. Once a leg has moved in its favour by a set multiple of the risk it was opened with, its desk trails the stop behind the best price — tighten only, budgeted per day, never a market close. Backtests do not see the trail, so that gap is measured live.
• The operator’s hand is respected. A hand close cancels the desk’s pending entries on that instrument and makes it stand aside; a model cycle already in flight cannot reopen the position. Only a losing stop-out starts a cooldown.
• Every desk names the market it trades — and is held to it. Per instrument it states trending up, trending down or ranging, beside a rule-based read of the same data. Code refuses an open against that statement, into a flat range or without a current call; exits are never checked. The operator can set a binding type for a few days, the managed legs follow it, and agent, rule and operator calls are each scored nightly.
• The operator’s view has a shelf life. A structured, expiring view — direction, conviction, horizon and levels — that every desk must answer each cycle and that is scored when it expires. It advises the model; it never gates the executor.
• Desks of the same class compare notes across exchanges. Each global commodity desk reads its China counterpart’s current view every cycle — COMEX gold beside Shanghai gold, NYMEX crude beside Shanghai crude, CBOT corn beside Dalian corn, US notes beside China’s treasury futures — and lessons that survive one desk’s review are copied to the other under caps.

The improvement loop — propose, replay, gate, promote

An agent improves the same way a strategy does: by beating the incumbent on evidence it could not have seen.

• One change at a time. A challenger differs from the champion in its prompt or in its memory setting — never both.
• Replayed on history, month by month, with incomplete windows marked unusable rather than scored.
• Only windows after the model’s own knowledge cut-off may decide. Everything earlier is directional evidence.
• To be promoted it must win most of those windows, earn a positive expectancy, not deepen drawdown, clear zero on a bootstrap interval, and not get sloppier. Promotion carries a cooldown and a one-action rollback.

8.3  The research agent — the loop that rebuilds the book

The desks manage risk inside a portfolio that already exists. The research agent runs the machinery that decides what the portfolio should be — one on each research box, at operator authority, doing everything a desk quant would do except the single act reserved for a person: promoting a portfolio into production.

• It keeps the search space honest. A signal library grows monotonically unless something prunes it; ours had grown roughly twentyfold from its original set, until a full optimiser sweep no longer fitted the cadence it was meant to run on. The agent holds the active pool to a compute budget, retiring on measured evidence in a fixed order — live results first, then optimiser results, then how often a signal is actually used.
• It sets the campaigns and runs them, instead of an operator hand-picking a scoring objective per strategy family.
• It measures live against backtest, leg by leg and signal by signal. Each leg is judged against its own live window rather than the full backtest curve — a decade of simulation against ninety days of trading measures the regime, not the strategy — and a leg with too few closed trades returns no verdict at all rather than a number computed from noise. Positions a human closed, and anything the circuit breaker flattened, are excluded from a strategy’s exit statistics: those were the operator’s decisions, not the model’s.
• It proposes against a stated objective, not a hand-picked construction method: candidates ranked on average daily P&L against the live book, among builds that beat the live book’s own out-of-sample result and stay inside its drawdown and margin envelope. A guard that cannot be evaluated counts as failed, never as passed.
• It proposes the capital ramp; a person approves it. The live book grows in steps tied to NAV milestones, each with its own construction recipe, from growth towards preservation. Every Saturday the agent checks the next step’s gates against the circuit breaker’s own daily NAV record and, on a pass, proposes a step built under the next recipe. Nothing switches a step on its own, in either direction, and a missing input reports the step blocked, never assumed. The book is in its first step today.
• Code it writes is reviewed before it can influence anything. Deterministic checks — static analysis, the known look-ahead traps, and a replay probe that truncates history to catch a module reading bars it should not see — plus a separate model reviewer. Both must pass, and the module then goes through the same promotion pipeline as a human-written one.

The desks may act on risk inside limits they cannot argue past; the research agent may change what gets traded only by convincing a person.

10 · Data Terminal

The macro cockpit.

The Data Terminal is where the outside world enters the decision flow. It gathers the macro, market, news and positioning data no single strategy looks at, turns it into a picture a person can read in a minute, and hands the same picture to the investment committee — whose verdict reaches the strategies and the agent desks as a small, machine-readable bias. Strategies never read the internet and desks never scrape a website.

What it gathers

• A fast tier, every minute — index levels, FX, ten-year yields and international commodity benchmarks for every market the firm follows.
• A slow tier — inflation, unemployment, growth, purchasing-manager indices and policy rates; official sources first, a second source only to fill a blank.
• The onshore China tape — every China commodity contract the books trade, beside its international counterpart and the premium between them.
• Sectors across six regions — China, US, Japan, Hong Kong, Europe and Korea on one sector spine, so a rotation reads as global or local at a glance.
• China liquidity, positioning and curves — funding rates, cross-border flows, margin balances, index-futures positioning and the rates curves.
• An FX tier — the pairs and the dollar index, two-year yields and policy rates, how leveraged funds are positioned against their three-year range, the gap between the yuan fixing and the market, and the slow drivers a currency desk watches: disaster and hurricane alerts, the El Niño cycle, European gas storage. Each source is dropped, never guessed, once it is past its age limit.
• Calendar and news — an impact-ranked calendar with actual against forecast, and news in English and Chinese, translated and tagged by a model for the markets — and the currencies — it touches.

What it produces

• The brief and the regime call — the drivers that matter and a regime label per region, refreshed with every committee sitting.
• A committee before every open — nine sittings each weekday, about half an hour before each market open in the live books, from the Tokyo morning through the London FX open to the Chicago evening re-open. Each produces a per-market stance, a stance per China commodity group and per global commodity class — shown side by side as one China | Global commodities block — a sector view and a view on each of eight currencies, with every role’s report archived.
• A bias that cannot strand a position — pushed to both clusters within minutes, clamped on arrival, ignored once stale, and applied to new entries only: exits, stops and targets never see it.
• A scorecard — every stance graded against the market’s next move, by market and by region, so the layer earns its weight.
• Alerts and Ask — regime changes and data surprises flagged, and questions answered from the terminal’s own data.

Isolated by design. A separate, read-only process on the research cluster with no write path into trading; a failed source keeps its last good value or shows a gap, never a guess; the only thing the trading side consumes is a bias file that is either fresh or ignored. On the web and as a mobile app, behind per-user grants and two-factor login.

12 · Operations and infrastructure

A supervised fleet.

Every process has a starter, a watchdog and somewhere to report.

• Two clusters, three boxes each — market data, trading, and research — laid out identically in both regions. The Global research box also hosts the two paper books, on their own market data, order manager and portfolio manager, so nothing paper shares a process with the live book.
• A broker gateway with no desktop behind it. The live book’s gateway runs headless and logs in with a passkey, re-authenticates itself on schedule, and is probed every minute for a session that has silently died or turned read-only — either is alerted and recovered automatically.
• Supervision. Per-minute watchdogs revive strategies, the circuit breaker, the agents and the feeds; the desk starts and stops any of it from the ops channel.
• Deploys. Every box runs a named commit on its own branch, promoted by a release tool that prints a restart plan, refuses a box carrying uncommitted changes and never restarts anything itself. The gate is a full test run compared against a recorded baseline. Nothing is copied to a production box by hand.
• One terminal, desktop and phone — positions and their stops, orders with algorithm progress, a watchlist with an FX tab, the risk panel, health and attribution, the roll calendar, the committee, the agents with their market-type and operator-view cards, and one view of every position across all five books with per-book return auto-close.
• A read-only research app for studying candidates and agents from a phone.

Appendix · Facts and figures

The platform in numbers.

All Sharpe and drawdown figures are backtest or validation statistics computed on FX-converted NAV in the book’s base currency, with the circuit breaker applied where stated. Live performance records are available to qualified investors on request.