Howie Huang

An all-inclusive Bitcoin blockchain analytics platform with automated parsing, graph construction, and ML experimentation.

ChainWeave ingests raw Bitcoin blockchain data, transforms it into behavioral subgraphs via a distributed ETL pipeline, and classifies those subgraphs using 23 graph neural network architectures — enabling researchers and analysts to identify patterns like illicit exchanges, darknet activity, and mixing services at scale.

Vision

A future where every blockchain transaction is transparent, every illicit pattern is detectable, and graph intelligence bridges the gap between raw on-chain data and actionable insight — empowering enterprises to protect financial systems and researchers to advance the science of graph analytics.

Mission

ChainWeave transforms open-source blockchain data into structured graph representations and applies state-of-the-art graph neural networks to uncover behavioral patterns at scale. The project serves two complementary missions:

• Research. Advance the frontier of graph machine learning on real-world financial networks by publishing reproducible studies, benchmarking novel GNN architectures against established baselines, and releasing open datasets and tooling to the academic community.
• Application. Deliver production-grade graph AI capabilities for fraud detection, anti-money laundering (AML), sanctions compliance, and risk scoring — helping financial institutions, exchanges, and regulators identify illicit activity faster and with greater precision than traditional rule-based systems.

Why ChainWeave?

Blockchain analytics today is fragmented: parsing tools, graph databases, and ML frameworks live in separate silos, requiring brittle glue code and manual hand-offs. ChainWeave unifies the entire pipeline — from raw blk*.dat files to interactive experiment dashboards — in a single, reproducible platform.

Key Capabilities

• Automated blockchain parsing — Ingest Bitcoin block files in parallel via Celery workers, writing parsed transactions, inputs, outputs, and SegWit witness data directly to columnar Parquet without an intermediate document store.
• Distributed ETL — A 4-stage PySpark pipeline performs UTXO resolution, distributed entity-cluster discovery via GraphFrames connected components, multi-source label enrichment, and node-feature assembly into ML-ready PyTorch Geometric tensors.
• 23 GNN architectures — From classic (GCN, GIN, GAT) to state-of-the-art (GraphGPS, DiffPool, PNA, SAT), all accessible through a single model registry.
• 8 swappable readout heads — Subgraph-level pooling strategies (sum, mean, max, attention, sort, Set2Set, GMT, GIN-style) for graph-level classification on entity-resolved clusters.
• Downstream ensembles — Plug XGBoost or Random Forest classifiers on top of GNN embeddings for additional classification power.
• Full-stack web interface — React dashboard for monitoring infrastructure, launching experiments, and visualizing graph structures.
• Async job orchestration — Celery workers handle long-running parsing, ETL, training, and inference jobs with real-time WebSocket progress updates.

Architecture

A separate text-mining sidecar runs in parallel with the four ETL stages, extracting OP_RETURN, input, witness, and coinbase script text and joining it to wallet labels for behavior-stratified text analysis.

ETL Pipeline

The PySpark / DuckDB ETL transforms raw blockchain data into ML-ready graph tensors in four stages:

GNN Model Zoo

ChainWeave includes 23 graph neural network architectures organized into four tiers, all accessible through a single build_model() factory:

Eight swappable graph-level readout heads — sum, mean, max, global attention, sort pool, Set2Set, GMT (Graph Multiset Transformer), and GIN-style concatenation — can be paired with any architecture, and downstream XGBoost or Random Forest classifiers can be trained on the resulting GNN embeddings.

Research Agenda

Tech Stack