| Machine Learning |
|
| Data Science |
|
| Big Data |
|
| Client-side |
|
| Server-side |
|
| Package Management |
|
| CLI & Markups |
|
Favorite Editors: Notepad++, VS Code
Agentic, fully automatic pipeline that batch-tests empirical win rate on ANY condition and ANY financial asset. Counts P(up | condition) − P(up) straight from history, differences the conditional CDF into a PDF to recover local edges, and reduces each family via shrinkage-weighted Naive Bayes.
Agent-oriented export pipeline for WeChat chat histories (Python + SQL). Supports both 1-on-1 and group conversations, and outputs LLM-readable compressed text. Table schema reverse-engineered independently.
Offline data warehouse for fraud detection on 6.3M PaySim transactions, built on PySpark + Hive + Hadoop. 3-layer architecture (ODS → DWD → ADS): DWD uses ORC + Snappy with derived fraud features; ADS outputs per-type fraud rates and a high-risk account registry.
Sequence-to-sequence 30-day forecasting on multi-feature on-chain data (PyTorch). Addresses lag-1 degeneracy via logarithmic detrending, orthogonal multi-feature inputs, and MambaSSM's selective state gating. Benchmarks six architectures — MambaSSM achieves best performance and strongest resistance to trivial state copying.
From-scratch replication of GPT-2 (162M params, 768-dim embeddings, 12 heads, 12 layers), based on rasbt/LLMs-from-scratch. Learned positional embeddings, pre-LayerNorm, GELU activation, causal masking, and BPE tokenization via tiktoken. Trained on the Harry Potter corpus with AdamW and a cross-entropy next-token objective.
ConvNet on MNIST/EMNIST; ResNet-34 and ResNet-101 on CIFAR-10/100. ResNet-34 uses basic blocks (two 3×3 convolutions), ResNet-101 uses bottleneck blocks (1×1 → 3×3 → 1×1). Core focus: how identity shortcut connections and network depth interact across datasets of varying difficulty.