Skip to content

RESUME.md

Latest commit

 

History

History
169 lines (115 loc) · 9.9 KB

RESUME.md

File metadata and controls

169 lines (115 loc) · 9.9 KB
Raw

📋 Professional Resume Entry

TRD-GNN: Temporal Graph Neural Networks for Bitcoin Fraud Detection

Independent Research Project | November 2024 - November 2025

For Technical Research Positions

Temporal Graph Neural Networks for Bitcoin Fraud Detection | Independent Research
November 2024 - November 2025

  • Developed zero-leakage temporal GNN for cryptocurrency fraud detection with 7/7 test validation, achieving 0.5846 PR-AUC on 203K Bitcoin transactions under strict temporal constraints
  • Conducted systematic 9-experiment investigation proving that simple architectures outperform complex models by 108% on small datasets through controlled ablation studies
  • Pioneered novel GNN-tabular fusion methodology achieving +33.5% performance improvement through heterogeneous graph embeddings combined with domain features
  • Quantified and reduced "temporal tax" from 16.5% to 12.6% (23.5% reduction) through architectural optimization, establishing new benchmark for honest temporal evaluation
  • Published 6 research contributions with complete failure-to-success documentation demonstrating scientific method and reproducible Kaggle implementation
  • Technologies: PyTorch Geometric, XGBoost, Python, CUDA | DOI: 10.5281/zenodo.17584452 | GitHub: BhaveshBytess/TRDGNN

For ML Engineering Positions

Bitcoin Fraud Detection System with Temporal Graph Neural Networks
Personal Project | Nov 2024 - Nov 2025

  • Built production-ready temporal GNN achieving 0.58 PR-AUC on 203K-node heterogeneous graph with strict zero-leakage temporal constraints
  • Engineered TRD sampler enforcing temporal ordering (time(neighbor) ≤ time(target)) with 100% test coverage (7/7 passing), preventing future information leakage in graph sampling
  • Optimized model architecture reducing parameters 10x (500K → 50K) while improving performance by 108%, enabling deployment on resource-constrained systems
  • Developed hybrid fusion system combining 64-dim graph embeddings with 93 tabular features for +33.5% performance gain on wallet-level fraud detection
  • Documented complete research cycle including failures, ablations, and solutions with fully reproducible Kaggle notebooks and comprehensive technical reports
  • Stack: PyTorch Geometric, XGBoost, scikit-learn, Python, pytest | GitHub: BhaveshBytess/TRDGNN | DOI: 10.5281/zenodo.17584452

For Data Science Positions

Graph-Based Fraud Detection Research | TRD-GNN Project
Nov 2024 - Nov 2025

  • Analyzed 203K Bitcoin transactions (182 features, 49 timesteps) using heterogeneous temporal GNNs, achieving 0.5846 PR-AUC under realistic temporal constraints
  • Performed systematic ablation experiments (9 total) isolating root causes of model failures, improving performance by 108% through principled architectural simplification
  • Created novel fusion methodology combining graph neural network embeddings with tabular features for +33.5% lift in wallet-level fraud detection
  • Developed comprehensive evaluation framework quantifying temporal evaluation costs ("temporal tax") and demonstrating 23.5% reduction through design optimizations
  • Published open-source implementation with full documentation, unit tests (7/7 passing), and reproducible experiments on Kaggle platform
  • Tools: PyTorch Geometric, XGBoost, Pandas, NumPy, Matplotlib, Seaborn, Jupyter | DOI: 10.5281/zenodo.17584452

Condensed (One-Line for Skills Section)

Temporal GNN Fraud Detection | Developed heterogeneous temporal GNN for Bitcoin fraud detection (0.58 PR-AUC, 203K nodes), achieving +108% improvement through systematic ablations and +33.5% fusion synergy | PyTorch Geometric, XGBoost | GitHub | DOI: 10.5281/zenodo.17584452

📊 Key Quantitative Metrics

Scale:

  • 203,769 transactions, 100,000 addresses, 421,985 edges
  • 182 features per transaction, 49 temporal timesteps
  • Heterogeneous graph: 4 edge types (tx→tx, addr→tx, tx→addr, addr→addr)

Performance:

  • Transaction-level: 0.5846 PR-AUC, 0.831 ROC-AUC (E7-A3 model)
  • Wallet-level: 0.3003 PR-AUC, 0.890 ROC-AUC (E9 fusion)

Improvements:

  • +108% improvement over initial complex model (E6: 0.281 → E7-A3: 0.585)
  • +4.7% improvement over homogeneous baseline (E3: 0.558 → E7-A3: 0.585)
  • +33.5% fusion synergy (tabular: 0.225 → fusion: 0.300)
  • 10x parameter reduction (500K → 50K) with performance gain

Efficiency:

  • Temporal tax reduced 23.5% (from 16.5% to 12.6%)
  • 7/7 unit tests passing for zero-leakage temporal sampler

Research Output:

  • 6 novel contributions with high citation value
  • 9 complete experiments (E1-E9) systematically documented
  • 4 technical reports with comprehensive methodology
  • 100% reproducible on Kaggle with preserved checkpoints

🎯 Skill Highlights

Technical Skills Demonstrated

  • Deep Learning: PyTorch, PyTorch Geometric (GNN architectures)
  • Machine Learning: XGBoost, scikit-learn, ensemble methods
  • Graph Theory: Heterogeneous graphs, temporal constraints, message passing
  • Data Engineering: Large-scale graph construction, feature engineering
  • Software Engineering: Unit testing (pytest), modular architecture, version control (Git)
  • Research Methods: Ablation studies, systematic investigation, failure analysis

Soft Skills Demonstrated

  • Scientific Rigor: Complete failure-to-success documentation (E6→E7→E9)
  • Problem Solving: Root cause analysis through systematic ablations
  • Communication: Comprehensive technical reports and documentation (2000+ lines)
  • Persistence: Recovered from 49.7% failure to achieve 108% improvement
  • Innovation: Novel fusion methodology with measurable 33.5% synergy

💼 LinkedIn Project Section

TRD-GNN: Temporal Graph Neural Networks for Fraud Detection

Developed first zero-leakage temporal GNN for cryptocurrency fraud detection, achieving 0.58 PR-AUC on 203K Bitcoin transactions. Conducted systematic 9-experiment investigation proving simple architectures outperform complex models by 108% on small datasets. Pioneered novel GNN-tabular fusion approach achieving +33.5% improvement. Published 6 research contributions with complete open-source implementation.

Technologies: PyTorch Geometric · XGBoost · Graph Neural Networks · Temporal Modeling · Python · CUDA

Repository: https://github.com/BhaveshBytess/TRDGNN
DOI: https://doi.org/10.5281/zenodo.17584452

🗣️ Elevator Pitch (30 seconds)

"I developed a temporal graph neural network system for Bitcoin fraud detection that properly handles time constraints—something most GNN research ignores. Through systematic experiments, I discovered that simpler architectures work better on small datasets, improving performance by 108%. I then created a novel fusion approach combining graph and tabular features, achieving an additional 33.5% improvement. The complete system is production-ready with full test coverage and open-sourced with 6 distinct research contributions."

📧 Email Signature Addition

Bhavesh Bytes
Machine Learning Researcher
GitHub: @BhaveshBytess
Research: Temporal GNNs for Fraud Detection | DOI: 10.5281/zenodo.17584452

📝 Cover Letter Snippet

"...My recent work on temporal graph neural networks demonstrates my ability to conduct rigorous research while building production-ready systems. When my initial heterogeneous GNN model failed by 49.7%, I didn't abandon the approach. Instead, I designed systematic ablation experiments to isolate the root cause, discovering that architectural complexity—not the heterogeneous structure—was the issue. By simplifying the architecture, I achieved a 108% improvement while reducing model size by 10x. This experience taught me the value of systematic investigation and the importance of documenting both failures and successes. The project resulted in 6 distinct research contributions and a novel fusion methodology achieving +33.5% improvement, all with full test coverage and comprehensive documentation ready for production deployment."

🎓 For Academic CV

Research Projects

Temporal Graph Neural Networks for Bitcoin Fraud Detection
Independent Research, November 2024 - November 2025

Systematic investigation of heterogeneous temporal GNNs for fraud detection addressing temporal leakage in existing research. Developed zero-leakage TRD sampler, conducted 9-experiment systematic study, and pioneered GNN-tabular fusion methodology. Key achievements: (1) Reduced temporal evaluation cost by 23.5%; (2) Proved Architecture > Scale principle (+108% improvement through simplification); (3) Achieved +33.5% fusion synergy; (4) Published 6 novel contributions with complete failure-to-success documentation. Production-ready implementation with 7/7 test coverage.

Publication-Ready Outputs:

  • Complete technical documentation (~2000 lines)
  • Reproducible Kaggle notebooks with all experiments
  • Open-source implementation (MIT license)
  • Comprehensive methodology reports
  • DOI: 10.5281/zenodo.17584452

Research Contributions:

  1. Zero-Leakage Temporal Sampler (production-ready, 7/7 tests)
  2. Temporal Tax Quantification & Reduction (16.5% → 12.6%)
  3. Architecture > Scale Principle (systematic proof via ablations)
  4. Successful Heterogeneous Temporal GNN (+4.7% over baseline)
  5. Architecture-Induced Collapse Discovery (failure mode identification)
  6. GNN-Tabular Fusion Synergy (+33.5% improvement)

Technologies: PyTorch Geometric, PyTorch, XGBoost, scikit-learn, Pandas, NumPy, Matplotlib, pytest, Git

Choose the format that best matches your target role!

For more details, see the full project at: https://github.com/BhaveshBytess/TRDGNN