Harrishkumar Loganathan

Architected graph-based identity fraud detection for onboarding scenarios, improving fraud-capture recall by 14% using PySpark, PyTorch Geometric (GraphSAGE), and Transformers. Built entity-resolution and velocity pipelines for reused/obscured attributes using SQL, Kafka, Spark Structured Streaming, and Redis to strengthen linked-identity detection. Developed multimodal document verification and synthetic/forged ID detection using OpenCV, OCR, Vision Transformers, PyTorch, and liveness techniques, improving detection by 12%. Implemented risk fusion with deepfake/camera-injection/device intelligence and served models via FastAPI using ONNX/TensorRT. Delivered real-time first-party fraud scoring for legitimate identities using XGBoost/LightGBM with Feast and PostgreSQL/Snowflake, and operationalized model governance using MLflow, Airflow, Kubernetes, SHAP, and Evidently for improved false-positive control (11%).

Led architecture for an enterprise support triage system using Python, scikit-learn, BERT/DistilBERT, and XGBoost/LightGBM, reducing manual reassignment by 18% and improving routing F1 by 11%. Directed NLP pipelines with PySpark, Pandas, spaCy, and Hugging Face Transformers to prepare 250K+ conversations for classification, sentiment, prioritization, and routing. Built and productionized a RAG customer-support copilot using Azure OpenAI, GPT-4, LangChain, sentence transformers, Elasticsearch, and FastAPI, reducing agent research time by 26% across workflows. Improved retrieval with BM25/vector/hybrid search and reranking, raising Recall@5 by 13% and standardizing relevance benchmarks. Implemented summarization and similar-ticket recommendations with GPT-3.5, PostgreSQL, and Redis, reducing response-preparation time by 21% while improving consistency. Established LLM evaluation and guardrails with MLflow, pytest, Recall@K, MRR, NDCG, groundedness, and hallucination checks to catch 17% m