Projects

From problem to result.

My personal labs — open-source or shipped. This is where I test my engineering standards outside the day job.

Selected work

Rubeau

Windows · AI agents

Windows app in testing, by invitation. No public release yet.

Visit rubeau.fr →

The problem: The office, rearranged. A company of AI agents you run without writing code, and watch at work.
The approach: Built on Claude, with your own API key or a local model.
The result: Windows app in testing, by invitation. No public release yet.

StackWindows · ClaudeVisit rubeau.fr →

Quant Validation Harness

Python · Statistics

Every strategy must clear this shield before going live — proof before promise.

View on GitHub →

The problem: Telling a real edge apart from noise and overfitting, before risking a single dollar.
The approach: A statistical validation harness: Deflated Sharpe Ratio, Combinatorial Purged CV, alpha/beta measures and a directional gate.
The result: Every strategy must clear this shield before going live — proof before promise.

StackPythonView on GitHub →

DataForge

MQL5 · Data integrity

Same symbol, three brokers: tick value at 0.1 against 1 — a factor of ten. And real_volume at zero across all three sources since 2019, contradictory before. Measured, not assumed.

View on GitHub →

The problem: You backtest and trade MT5 data without knowing what's really in it: brokers differ silently, and the volume on screen can be a mirage.
The approach: An instrument passport in every export header, a per-year volume table read via CopyRates (not the tester), and an MQL5 ↔ Python parity reference.
The result: Same symbol, three brokers: tick value at 0.1 against 1 — a factor of ten. And real_volume at zero across all three sources since 2019, contradictory before. Measured, not assumed.

StackMQL5View on GitHub →

HorizonProbe

MQL5 · Statistical measurement

The boundary made visible. And the rigour with it: a MQL5 → Python transcription announced 2.1e-10 when the real MQL5 output was 1.339 — ten orders of magnitude apart, caught before concluding.

View on GitHub →

The problem: Is your holding period profitable by construction? The toll is fixed, the move grows as the square root of time — so there's a horizon below which even a perfect direction loses.
The approach: Measured across every rolling window, with no Gaussian assumption — the break-even threshold located horizon by horizon, not an average.
The result: The boundary made visible. And the rigour with it: a MQL5 → Python transcription announced 2.1e-10 when the real MQL5 output was 1.339 — ten orders of magnitude apart, caught before concluding.

StackMQL5View on GitHub →

RiskCockpit

MQL5

A guardrail that watches risk in real time and protects traders from their own worst impulses.

View on GitHub →

The problem: Traders skip the news, oversize their lots and trade on tilt — discipline breaks under pressure.
The approach: An MQL5 risk & discipline indicator: news-gate, drawdown-budgeted lot sizing, tilt-lock. Its economic calendar comes from its companion service, RCNewsFeeder.
The result: A guardrail that watches risk in real time and protects traders from their own worst impulses.

StackMQL5View on GitHub →

RCNewsFeeder

MQL5 · Service

RiskCockpit classifies news against the prop-firm rules. The service stays read-only: it writes one file, and touches no trade.

View on GitHub →

The problem: MQL5 blocks network access from indicators: the request fails even with the URL whitelisted — that right belongs to services and Experts only. RiskCockpit being an indicator, it cannot fetch the economic calendar itself.
The approach: A companion MQL5 service — services are allowed on the network. It downloads the economic calendar every hour and writes a single file the indicator reads: a file bridge, with nothing changed inside the indicator.
The result: RiskCockpit classifies news against the prop-firm rules. The service stays read-only: it writes one file, and touches no trade.

StackMQL5 (service)View on GitHub →

StrategyDeck

MQL5 · Manual cockpit

Published on the MQL5 Market. Every order comes from a click, and when a measurement is missing the tool answers “not measured” instead of inventing a number.

View on the MQL5 Market →

The problem: In manual trading you decide without knowing the real round-trip cost of the symbol, the hit rate a target needs once that cost is counted, or the position size the account can actually hold.
The approach: A MetaTrader 5 cockpit that draws and measures, without ever deciding or sending an order on its own: round-trip costs read from the symbol, observed drawdown, holdable size, and a local AI co-pilot kept away from execution.
The result: Published on the MQL5 Market. Every order comes from a click, and when a measurement is missing the tool answers “not measured” instead of inventing a number.

StackMQL5View on the MQL5 Market →

le Journal

Rust · React · Python

Your data stays on your machine; the tool mirrors your own patterns back and trains your discipline.

Learn more →

The problem: Classic trading journals log trades but never actually fix the behavior.
The approach: A local-first desktop trading journal: a Rust core (Tauri 2), a React/TypeScript UI, a 100% local SQLite store and a Python sidecar for MetaTrader 5; a discipline-focused coaching engine.
The result: Your data stays on your machine; the tool mirrors your own patterns back and trains your discipline.

StackTauri 2 · React · SQLiteLearn more →

ProSessionBox

MQL5

A clean utility that traders keep pinned to their chart.

View on GitHub →

The problem: Knowing at a glance which market session is live and when liquidity arrives.
The approach: An MQL5 market-session tool: clear visualization of sessions, overlaps and key time windows.
The result: A clean utility that traders keep pinned to their chart.

StackMQL5View on GitHub →

The Solution Maker

Next.js · i18n

This very site — a living showcase of how I build: precise, fast, international.

Visit the site →

The problem: A recruiter portfolio AND a product store, in 3 languages, with no performance compromise.
The approach: Next.js 16, App Router, i18n FR/EN/FA, MQL5 store, static rendering and careful SEO.
The result: This very site — a living showcase of how I build: precise, fast, international.

StackNext.js 16Visit the site →