Arun Nathi

AI/ML Engineer with 4+ years building and shipping production ML systems in fintech, including real-time fraud detection, Retrieval-Augmented Generation (RAG), LLM inference optimization, and end-to-end MLOps on AWS for production-scale financial systems.

Deep expertise across the full ML stack: classical ML (XGBoost, LightGBM, CatBoost), deep learning (CNN, RNN, LSTM, BERT), GenAI (RAG, LangChain, GPT-4, LLaMA), and MLOps (SageMaker, Docker, Kubernetes, CI/CD). Track record of measurable impact across fraud detection, latency reduction, query automation, and compliance workflows, with SHAP explainability, A/B testing, and model risk governance practices.

AI/ML Engineer - PayPal - United States

(2024-11)

• Designed and deployed a RAG system over 100K+ internal financial documents using FAISS dense retrieval and LangChain; chose FAISS over Pinecone to cut managed-service latency and infra cost, reduced query resolution time by 60% across 10+ operational teams.
• Built GPT-4 summarization pipelines for compliance transaction reports, processing 500+ reports/week with structured output validation and human-in-the-loop review gates; eliminated 20+ hours/week of manual work and improved audit traceability across 3 compliance workflows.
• Developed transformer-based anomaly detection models for real-time transaction classification; initially used isolation forest (poor recall on rare fraud types), pivoted to fine-tuned BERT classifier, achieved 91% precision and flagged ~$200K in potential losses within 2 months of deployment.
• Engineered OCR + NER extraction pipeline for invoice processing; handled 3,500+ documents/month with auto-retry on low-confidence extractions, reduced manual errors by 35% and cut processing time by 50%.
• Optimized LLM inference on SageMaker via dynamic batching, FP16 quantization, and model co-location, reduced latency from 2.4s to 0.7s (71% improvement) with no accuracy regression.
• Owned full deployment lifecycle for internal AI dashboards (model API → monitoring → alerting) serving 500+ accounts daily; set up CloudWatch drift detection with automated retraining triggers to maintain production model quality.
• Designed inference infrastructure to handle concurrent query spikes of 200+ req/sec using SageMaker autoscaling endpoints and Redis caching for repeat queries, sustained p99 latency under 1s during peak compliance reporting windows.
• Established evaluation framework combining offline metrics (precision, recall, NDCG) with online A/B testing and user feedback loops (retrieval acceptance rate, correction signals) to continuously improve RAG quality across model iterations.
• Partnered with finance and product stakeholders to translate model outputs into actionable insights, supporting data-driven decision-making for operational and risk management strategies.

ML Engineer - LTIMindtree - India

(2021-10 - 2023-07)

• Built statistical and machine learning models (regression, classification, gradient boosting) for financial risk scoring over ~50M records/month across 5,000+ accounts; engineered 120+ features with automated validation, improving predictive accuracy and enabling real-time anomaly detection.
• Designed PyTorch CNN-based fraud detection models across 8 transaction types; addressed severe class imbalance (1:500 fraud ratio) using SMOTE + focal loss, improved F1 from 0.61 to 0.84 and prevented ~$1.2M in estimated annual fraud losses.
• Built and owned a Transformer-based NLP document classifier covering 15+ financial categories in production; reduced misclassification by 40% over the rule-based baseline and cut manual processing time by 40%, directly improving straight-through processing rates.
• Engineered XGBoost + LightGBM stacked credit risk models deployed across 8 loan portfolios; delivered 15% risk reduction and produced SHAP explainability reports adopted by risk and compliance leadership for model governance sign-off.
• Established end-to-end model validation framework, cross-validation, error analysis, and shadow deployment against statistical baselines, ensuring all models met production reliability thresholds before release.
• Designed and executed time-series backtesting and Monte Carlo simulation framework across 5 consecutive quarters to validate model stability under distribution shift; findings directly satisfied internal model risk governance requirements and informed production release decisions.

Junior ML Engineer - LTIMindtree - India

(2020-05 - 2021-09)

• Owned end-to-end development of customer churn classifiers (Random Forest, XGBoost, logistic regression) with GridSearchCV tuning; XGBoost model selected for production, reducing false positives by 150 cases/month and directly improving retention campaign targeting accuracy.
• Deployed sales forecasting models to production via Flask REST API serving 10 regional teams; achieved 1.2 RMSE improvement over baseline, reducing manual reporting effort by 25 hours/month and establishing the team's first automated forecasting system.
• Engineered automated model retraining pipeline triggered by data drift detection; reduced model degradation incidents by 60% and cut retraining time from 3 days to 4 hours, establishing the team's first MLOps practice.

M.S. - Data Engineering (Machine Learning & AI) - University of North Texas (2023 - 2025)