Pankaj Maurya

Most ML projects die in a notebook. Mine ship — and keep running. I'm Pankaj — a Senior ML Engineer with 5 years building systems that go from raw data to a live endpoint, not just a proof of concept. 3 years shipping production ML at scale, 2 years in analytics before that, so I understand both the model and the business question behind it. Proof, not promises: ✅ 0.92 ROC-AUC fraud detection model on 590K+ transactions with a ~3.5% fraud rate (LightGBM, time-based CV, zero data leakage) ✅ Deployment cycles cut from 2 weeks → 3 days after building MLflow-based MLOps pipelines with automated drift monitoring ✅ 90% SQL runtime reduction (45 min → 5 min) on 1M+ row datasets — saved 40+ engineering hours/month ✅ 2.2% revenue lift from a dynamic-pricing model, validated through backtesting against flat pricing ✅ 99.9% uptime serving real-time NLP predictions to 50+ global stakeholders via FastAPI + Docker ✅ 22–26% forecast improvement over baseline on a 30,490-series retail demand model (Walmart M5) What I build: 🔹 Production ML Pipelines — XGBoost, LightGBM, Random Forest, LSTM — deployed on AWS SageMaker in Docker containers, not left in a Jupyter notebook 🔹 GenAI & NLP Systems — fine-tuned DistilBERT for intent classification, sentence-embedding clustering for theme discovery, VADER sentiment scoring, RAG pipelines with LangChain + FAISS 🔹 Forecasting & Time-Series — demand forecasting with lag/rolling-window feature engineering, Tweedie-loss models for intermittent zero-inflated demand, SARIMA baselines 🔹 Fraud, Churn & Anomaly Detection — imbalanced-class modeling with SMOTE, threshold tuning aligned to real business costs, not just accuracy scores 🔹 Explainability Built In — SHAP and Grad-CAM on every model, so your team can see why a prediction was made, not just trust a black box 🔹 Analytics & BI — Power BI dashboards, advanced SQL (window functions, CTEs), ETL pipelines cutting data prep time by 60% How I work: I run monthly A/B tests and hypothesis-driven experiments to validate that a model actually moves the metric it was built for — not just performs well offline. I treat every project like it's going into production, because for most of my career, it has. Whether you need your first fraud model, a forecasting system that survives real-world data drift, or a full MLOps setup so your team stops babysitting deployments — I'll scope it around the business outcome first, the architecture second. 📩 Let's talk about what you're trying to solve — I typically respond within a few hours.