Live

Running deployments, not screenshots of things that once worked. Each rebuilds itself from the code — so if a pipeline breaks, the page breaks with it.

• 

  SMB Growth Intelligence

  GA4 events through layered ETL, a point-in-time feature store, a propensity model and a data-quality engine. CI runs the whole pipeline on every push and publishes whatever it produced — the wall time on the page changes run to run because it is real.

  duckdb feature store bigquery-portable

  → launch source
• 

  Forecasting Baselines

  Naive, moving average, Ridge, XGBoost and an LSTM benchmarked walk-forward on sixteen years of real market data. None of them beat “tomorrow = today” on RMSE, which is the correct result — and the headline says so rather than burying it.

  walk-forward leakage-safe negative result

  → launch source
• 

  Geometry Dash RL

  DQN, PPO and a genetic algorithm trained in a headless simulator running four thousand times faster than real time, then dropped into the retail game through a C++ mod to clear a full official level.

  reinforcement learning sim-to-real pytorch

  → launch source