Skip to content
View PNX89's full-sized avatar

Block or report PNX89

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
PNX89/README.md

Quelin Zammit

The map is at pnx89.github.io: what each of these repositories compares against what, and why that is the same move every time.

Systematic trading developer. Malta, EU. I build the research and execution machinery behind quantitative strategies, and the automation that makes the results worth trusting.

That second half is the part I care most about. A strategy is a claim, and most of the work is deciding whether the claim survives contact with transaction costs, with the number of configurations you actually tried, and with a market regime you did not fit on. The infrastructure that answers those questions honestly is, to me, the interesting engineering.

Most of what I have built is private, either commercial or under NDA: a multi-agent research platform running an anti-overfitting funnel across a cloned MetaTrader farm, a live expert advisor with a custom fitness engine, and sixteen years of payments and back office systems at QPARTZ, the company I founded and am now in the process of stepping out of. The repositories below are the parts that can be shown, generalised onto synthetic data. Each is built to be checked rather than admired: no test needs an API key or a market data feed, the numbers in every README come out of a real run, and every limitations section says plainly what the tool cannot do.

Quantitative research

QUACKZ grades a backtest on the ways backtests overstate themselves: a deflated Sharpe against the number of trials you actually ran, the noise floor a search that size expects from pure chance, the basis points of cost at which the edge disappears, and a stationary bootstrap showing how much of the result is sampling luck. It implements Bailey and Lopez de Prado properly, worked examples pinned as tests. The demo searches two hundred random signals, keeps the best, and then fails it.

QUOTEZ hands MetaTrader 5 market data to an LLM agent as typed, read only tools over the Model Context Protocol. There is no order placement code anywhere in it, which is a structural property rather than a configuration flag: an agent cannot trade through an interface that has no path to an order.

Reliability, where automation meets money

QUIDZ is the money path: signature verification for two webhook schemes that disagree about replay defence, an idempotent ledger that survives duplicate and out of order delivery, and a reconciliation pass that classifies drift and closes a gate on outbound money movement rather than on ingest.

QUESTZ is about the automation failure that actually costs money: not the page that disappears, but the page that quietly moves a column while the job keeps logging in, keeps writing a CSV and keeps exiting zero with wrong numbers. It fails closed on structural drift before any bad data is written.

QUELLZ attacks a tool-using agent with a catalogue of prompt injections, then measures what a least privilege containment layer actually stops. It reports attack success rate and task utility together, because a defence that blocks everything scores perfectly on the first number and is useless.

Background

I have run production systems since 2010. Payments came first from the platform side, at a card payments processor in 2015, and then from the other side entirely, owning a live pipeline end to end for years afterwards. The systematic trading work has run alongside it throughout, on my own capital.

One theme connects all of it, and every repository here: designing for the failure that does not announce itself. The duplicate webhook. The backtest that only looks good because it was picked from two hundred. The page that moves a column without erroring. Systems that fail loudly are a solved problem; the expensive ones keep returning success.

Reachable on LinkedIn.

Pinned Loading

  1. QUACKZ QUACKZ Public

    Audits a trading backtest for the ways backtests overstate themselves: deflated Sharpe, noise floor, cost sensitivity, resampling, arithmetic shown

    Python

  2. QUOTEZ QUOTEZ Public

    MCP server exposing MetaTrader 5 market data to LLM agents as typed, read only tools. No order placement code exists in it

    Python

  3. QUIZZ QUIZZ Public

    Prices get revised: an as-of agent scored on right answers and on correct refusals, behind a gate whose pass mark rises with every variant tried

    Python

  4. QUELLZ QUELLZ Public

    Prompt injection red team and containment harness for tool using LLM agents. Reports attack success rate and task utility together

    Python

  5. QUIDZ QUIDZ Public

    Payment webhook reconciliation: signature verification for two schemes, replay defence, an idempotent ledger, a fail closed gate on outbound money

    Python

  6. QUESTZ QUESTZ Public

    Browser automation reliability: fails closed on structural page drift before bad data is written, with a circuit breaker and an evidence journal

    Python