The first commit was December 9, 2025. One line in the message: "REAL QuantDesk V4 — FULL PROJECT UPLOAD."
That "REAL" is doing a lot of work. There'd been earlier versions. Reports, training scripts, GPU runs that took hours overnight. But nothing that actually traded. This was the first time the whole thing existed in one place with a clear intention.
What was in it
The stack at that point: a momentum strategy with XGBoost as the signal engine, walk-forward backtesting validation, two shell scripts called auto_train.sh and overnight_train.sh, and a command-line tool called quantcli.py. That was it. ETH/USD only. Everything else was locked out.
The decision to hard-lock to a single pair at launch was deliberate. If you're learning to drive you don't start on a motorway. ETH/USD has deep liquidity, tight spreads, and it behaves like a proper market without the chaos of smaller coins. The XGBoost model had been trained on historical candles, validated with purged k-fold cross-validation so future data couldn't leak backward into the training set — a mistake that makes backtests look great and live performance catastrophic.
What it looked like
Looking back at those first files, the shape of everything that came after is already there. The configs were in YAML. The backtesting engine had an interface for strategies. There was already a test file called test_risk_engine.py.
I didn't know what I was building yet. I knew I wanted a system that could run unsupervised, size positions intelligently, and not blow up. At that point, "not blowing up" felt like a reasonable first target.
The trading desk didn't go live immediately after that commit. There was more work — the data truth gate, the crash recovery stack, weeks of watching paper results. But December 9 was the day the project existed.
Everything since has been refinement, automation, and the occasional disaster.