Abhinay Krishna Bodi

       

I build AI systems and the data infrastructure behind them. At Mastercard I trained fraud and ranking models on 4 TB+ of transaction data. At Brane I shipped sub-10ms WebSocket backends serving 5,000+ concurrent users and Spark forecasting pipelines. At NYU I research 4-bit LLM efficiency and teach NLP to 90+ students. Models that ship, systems that stay up.

• Mastercard | Data Scientist Intern

• New York University | Research Scholar

• New York University | Teaching Assistant, Natural Language Processing

• Brane Enterprises | Data Scientist

• Brane Enterprises | Software Engineer

• New York University (NYU Courant) | M.S. in Data Science (GPA: 3.8 / 4.0)
• VNR VJIET | B.Tech in Computer Science & Engineering (GPA: 3.9 / 4.0)

• llama.cpp | Merged PR #26536
  • Eliminated redundant audio-encoder chunks for short inputs by replacing 31-second preprocessing padding with the exact 201-sample FFT reflection-padding boundary, halving encoder chunks for affected inputs.

• VIP-MINGLE: Multimodal Interaction Corpus | Accepted at INTERSPEECH 2026
  • Co-authored a multimodal corpus built on transformer-based ASR, extracting aligned audiovisual cues and identifying statistically significant differences in turn-taking and participant enjoyment (p=0.037).

• satavahana Rust Tokio HFT

  • High-frequency options trading engine in async Rust. Streams 1.5GB+/4min tick feeds into lock-free DashMap storage with CPU-pinned workers, an 8-strategy signal pipeline, and Half-Kelly risk allocation.

• creditRisk_Prediction_GNNs PyTorch Geometric FA-GNN LSTM

  • Feature Attention Graph Neural Network with LSTM temporal modeling over dynamic borrower graphs, beating Random Forest and XGBoost baselines at 0.7707 ROC-AUC. Cut data-loading time 94.8% with Polars and explained feature-level risk with SHAP.

• cheque_forensics PyTorch Qwen3-VL NVIDIA DGX

  • Two-stage VLM document forgery pipeline combining a 655.8M C-RADIOv4-H backbone with 30B Qwen3-VL verification, trained on NVIDIA DGX Spark with a 13.2x training speedup at 0.89 precision.

• adaptive-pairwise-preferences Python Bayesian ML Active Learning

  • Bayesian latent factor model for sequential pairwise active feedback on the Netflix Prize dataset (100M+ ratings), cutting required feedback queries by 40% via information-gain maximization.

• ArcPay-Agentic-Financial-System Python Multi-Agent USDC

  • Agent-driven financial execution translating natural language to on-chain USDC payments and equities trading, guarded by a deterministic GuardianAgent liquidity and whitelist risk layer.

• H-ARC_challenge Python LLM Reasoning Program Synthesis

  • Tackles the ARC AGI benchmark through closed-loop hypothesis generation and neural code synthesis with 32B Qwen 2.5 Coder, reaching 11% accuracy against 0% for direct prompting.

• Languages: Python, SQL, R, Rust, C++, TypeScript, Scala, Java, Bash
• AI & ML: PyTorch, TensorFlow, XGBoost, Scikit-learn, SHAP, FAISS, vLLM, LangChain, LangGraph, RAG
• Data Engineering: Spark, PySpark, Kafka, Airflow, PostgreSQL, MongoDB, BigQuery, Snowflake, ClickHouse, Polars, HDFS
• Cloud & DevOps: AWS, GCP, Azure, Docker, Kubernetes, Jenkins, Linux, Git, CI/CD
• Analytics & Visualization: Tableau, Power BI, Excel/VBA, Matplotlib, Seaborn, A/B Testing