Back to all work
— Project 08
ML and forecasting

Rug Pull Detector

Crypto time-series ML

A machine-learning system for flagging meme-coin rug pulls from raw chart data. The important choice was a fair tournament: every model trains on the same feature window, so deep-learning and boosted baselines can be compared honestly.

Role
ML Engineer
Period
2025 to 2026
Status
Production
TSMixerLSTMTransformerXGBoostLightGBMPyWaveletsOptuna
— Chapter 01
System shape

How the system fits together.

Click any block to read about it
Fig. 01 — Rug Pull Detector architecture
— Chapter 02
Decisions and outcomes

The calls that shaped it.

  1. 01

    Built a multi-model tournament: TSMixer (winning architecture), LSTM, Transformer, XGBoost, and LightGBM all train on the same prepared windows so each model is evaluated against an identical benchmark. Stratified 5-fold CV with class weighting handles the heavy class imbalance.

  2. 02

    Designed the 50+ feature pipeline that downsamples raw 1-second ticks to 5-second buckets, then layers price velocity, RSI, ROC, buy/sell pressure, three-level wavelet decomposition, rolling-window stats (mean, std, Bollinger bands), and holder-concentration metrics from on-chain data.

  3. 03

    Production training loop: GPU with mixed precision, early stopping with patience, learning-rate scheduling, model checkpointing, and per-fold artefacts so any fold can be inspected after the run. Outputs full classification report with PR curve and confusion matrix.

— Aside
The interesting work isn't the stack. It's the boundaries.
— Chapter 03
How it runs

What it runs on.

  • 01
    PyTorch (TSMixer, LSTM, Transformer) plus XGBoost and LightGBM baselines
  • 02
    Stratified 5-fold cross-validation with class weighting
  • 03
    PyWavelets for multi-level wavelet decomposition features
  • 04
    Optuna hyperparameter search across all model families
  • 05
    CSV + JSON ingest with downsampling, outlier handling, and RobustScaler