• 🥇 I'm proficient with:

  • torch, scikit-learn, scikit-optimize, gplearn, numpy/scipy/statsmodels, pandas, xgboost/catboost/lgbm, matplotlib/seaborn/plotly, albumentations, category_encoders, HF transformers/diffusers, llamacpp, qdrant, fastapi, celery, streamlit

• 🥈 I know my way around:

  • spark, SQL, keras/tensorflow, OpenGL, OpenCV, CI/CD pipelines

• 🎖 Since I started, I've learned how to:

  • never accept the null hypothesis;
  • stop misinterpreting p-values;
  • read impurity-based, permutation and SHAP feature importances properly;
  • spot the outliers in several dimensions at once with Mahalanobis distance;
  • combat skewness with QuantileTransformer/PowerTransformer;
  • calibrate;
  • avoid the PCA trap in classification;
  • engineer features with symbolic regression;
  • effectively finetune SD using Dreambooth technique;
  • use vector databases to supply context for language models;

• 🎖 In the past year, I've learned:

  • the actual difference between Adam and AdamW;
  • proper separation of evaluation and decision making;
  • the power of generalized linear models;
  • numerous applications IoU may have;
  • the match between kernel regression and two-layer neural networks;
  • augmenting the embeddings rather than raw data;
  • that the LOOCV MSE for KNN equals the training MSE for (K+1)NN;
  • the importance of embeddings' distances for the reasoning of LLMs;
  • the difficulties of separating epistemic and aleatoric uncertainty;
  • how to evaluate ensembling quality using information theory;
  • how to beat OpenAI using a 7B model running on CPU.

Stackoverflow | DS Stackexchange | Cross Validated

Checkio