For quants
Most of the work in a quantitative workflow is not the idea. It is the capture pipeline, the order lifecycle, the reconciliation, and the long argument with yourself about whether a result is real. This is that scaffolding, already built and already hardened against the specific ways each part fails in production.
It supplies no edge, and this page makes no claim about what any strategy earns. What it offers is that the measurement is trustworthy and the execution is boring — which is the part that usually costs a team months.
What it costs to run
One small virtual machine. The whole stack — order book, detector suite, scoring, risk, execution and the web terminal — was built and measured on a 2-vCPU instance with 945 MiB of RAM. One instrument runs comfortably there; three or four is the measured ceiling rather than a guess.
- No market-data vendor. Book, trades, mark price and funding come from the exchange's own public streams and endpoints.
- No database to operate. State is flat files; telemetry is a single JSON snapshot written atomically.
- No native dependency in the web tier. The interface is pure Python, so it stays importable when a compiled wheel misbehaves.
- Your exchange account, your keys. Self-hosted. We never hold your credentials and never take custody.
Licensing is a flat subscription, not a share of anything you make, and not a per-seat charge. See plans.
Quantitative researchers
The expensive part of research is not running a regression. It is obtaining data you can trust and then not fooling yourself with it.
- Capture is already wired. Per-instrument tick files with the book, the signal values and the regime label recorded alongside the price, so a study reads one file instead of reconstructing state.
- Studies are scripts with verdicts, not notebooks. Each answers one question, prints the diagnostics that would invalidate its own answer, and refuses to return a verdict below a minimum sample — so a thin result cannot quietly become a conclusion.
- Negative results are kept as tests. Where a component or a signal was measured and found not to help, that finding is pinned by a test and recorded beside the code. A dead end stays closed instead of being rediscovered next quarter.
- Replay against recorded fixtures with ground-truth labels, so detector changes are compared on identical input.
The part that is hardest to buy is the catalogue of ways this measurement goes wrong, written down with the wrong answer each one produced. Pooling windows with different base rates once reversed a result's sign at high confidence. Counting an unchanged price as a directional win manufactured an effect out of nothing — the mid is unchanged on 86.8% of one-second intervals on a venue quantised at a cent. A rotated log file truncated history and flipped the sign of an answer. A frozen capture read as a quiet market until liveness was judged on top-of-book quantities rather than price, because a price can legitimately sit still and sizes cannot.
Quantitative developers
Sending an order is a morning's work. Everything after it is the year. This is the part that only appears once real money is moving.
- Order placement. Parent orders sliced with post-only prices re-clamped against the current touch, bounded by whether the signal that opened them still holds, and with unfilled slices retracted when it stops — because a passive order abandoned on the book can fill long afterwards, on the wrong side of a position you have since taken.
- Protective orders as a lifecycle, not a call. Coverage verified against quantity, so a position that fills in pieces cannot sit mostly unprotected while a check reports healthy. One owner may cancel. Creation claims its work before doing it, so concurrent fills cannot each place a pair.
- Reconnection that recovers. Listen-key keepalive, time re-sync and open-order reconciliation on every reconnect, so fills that landed during the gap are picked up rather than lost.
- Account-level resources shared, not multiplied. Rate limits and bans are per-account, so N instruments run over one request budget, one rate-limit breaker, one daily loss limit and one risk budget that divides as instruments are added.
- Protection survives your process. Stops and targets live on the exchange. One fired and closed its position 36 hours after the software had stopped running, which is the property any unattended run depends on.
Interfaces are declared per module and the mocks in the test suite are typed against those declarations, so a changed signature fails at the boundary. Around 1,278 tests, and the tests are themselves mutation-checked: a test that cannot fail when the behaviour it describes breaks is treated as a defect rather than as coverage.
Quantitative traders
After the fact, the only question is what actually happened. Most journals cannot answer it because they record the plan and not the outcome.
- Every row carries its own provenance. Realised P&L is recorded with a field saying whether it came from the exchange or from a local reconstruction, so a reader never has to guess which.
- Planned and realised geometry side by side. The stop committed at entry is kept next to the stop the exit actually hit, and the intended reward-to-risk next to what the trade delivered. The gap between those two is the single most useful number in the file, and it only exists if both are written down.
- Fees as charged, not as modelled. Commission is taken from the exchange's own per-fill figures across both legs, so net is net.
- Reconciliation against the exchange. The journal is checked against account income rather than trusted, and a disagreement is reported instead of absorbed.
- Bounded downside, with the direction chosen deliberately. Daily loss limit, throttling, cooldowns and a trail that cannot arm on a losing position. Each gate documents whether it fails open or closed and why, and each has a named environment switch so it can be turned off without a code change.
The terminal that shows all of this runs as a separate process that imports nothing from the trading stack and reads a published snapshot. It therefore holds no exchange credentials, has no code path that can place an order, and keeps serving last-known state while the runner is down.
What you still have to bring
Worth saying plainly, because the opposite is what most infrastructure pitches imply.
- The edge. This is execution and measurement infrastructure. It does not supply a predictive signal, and no component of it should be read as one.
- Your own account and keys, and the operational responsibility that comes with them.
- Your own conclusions. The study harness is built to make a weak result visible as a weak result. It will as readily tell you an idea does not work, and on this codebase it frequently has.
If that division of labour is the one you want, request access and we will send the agreement, the system requirements and an invoice.
Important information
Trading leveraged derivatives carries a substantial risk of loss and is not suitable for everyone. This page is not financial advice, nor an offer, a solicitation or a recommendation to buy or sell any instrument, nor a recommendation of any trading strategy. No performance claim is made anywhere on this site, and no past or projected result should be inferred from any description of the software.
We provide software and infrastructure. We do not manage client money, take custody of client assets, or exercise discretion over client accounts. Any prospective arrangement is subject to written agreement and to the licensing and suitability requirements of the relevant jurisdiction.