Fayaz – PyTorch, RAG, MLOps

An enterprise document intelligence platform that lets business users query large unstructured document repositories via natural language. The system combines RAG with hybrid retrieval — dense vector search, BM25, and cross-encoder reranking — for precise, context-grounded answers at scale. A separate fine-tuning pipeline adapts open-source foundation models (LLaMA 3, Mistral 7B) to domain-specific classification tasks, with full lifecycle management from dataset curation through production serving.

• Lead the document intelligence platform, driving the end-to-end architecture from retrieval layer to inference infrastructure.
• Built a hybrid retrieval pipeline combining dense vector search, BM25 scoring, and cross-encoder reranking to maximize answer precision on enterprise document corpora.
• Designed and implemented the LLM fine-tuning pipeline using LoRA adapters; made it fully config-driven so new experiments required no code changes.
• Covered the full fine-tuning lifecycle: dataset curation, training with Weights & Biases tracking, evaluation, and production serving via FastAPI on AWS SageMaker.
• Own the Kubernetes-based inference serving infrastructure, including the observability pipeline that monitors response quality and flags production anomalies.
• Built the standardized evaluation framework that gates every model version before promotion to production.

An ML-powered underwriting and actuarial risk assessment system processing applicant data, medical history, and third-party signals to produce risk tier scores that inform life insurance underwriting decisions. The platform served as the primary decision-support layer for underwriting teams, with strict regulatory oversight requiring documented model validation and explainability.

• Owned the machine learning systems used in underwriting and actuarial risk assessment, including the primary risk scoring model in production.
• Processed applicant data, medical records, and third-party signals to generate risk tiers and supporting documentation for underwriting decisions.
• Implemented SHAP-based model explanations across all production models to satisfy regulatory requirements; performed model validation for regulatory filings including out-of-time testing and fairness analysis.
• Built the continuous model monitoring and automated retraining pipeline.
• Redesigned the CI/CD deployment pipeline and rewrote the feature engineering layer into a versioned, tested Python library.
• Collaborated directly with actuaries and underwriting leadership to align model outputs with business requirements.

A fraud detection and predictive modeling platform for property and casualty insurance carriers, screening insurance claims using structured claim data, third-party signals, and behavioral patterns to surface high-risk cases for investigator review. The system reduced manual review load and improved fraud identification accuracy through a continuously improving feedback loop.

• Built and maintained fraud detection ensemble models that screened insurance claims using structured data, geospatial signals, and behavioral patterns.
• Implemented a feedback loop where fraud investigator flags on false positives fed back into training data, continuously improving model precision.
• Built the PySpark data pipeline handling data ingestion, quality checks, and feature transformation at scale.
• Handled model validation for regulatory filings and worked with the product team to integrate model scores into the claims management platform.
• Spent time with fraud investigators to understand how they used model outputs, informing feature design decisions.

A packaged machine learning product suite deployed across enterprise clients in telecom, retail, and financial services. Products covered churn prediction, customer segmentation, and demand forecasting, with delivery managed end-to-end from data ingestion through production deployment.

• Deployed and customized packaged ML products across enterprise clients in telecom, retail, and financial services.
• Delivered churn prediction, customer segmentation, and demand forecasting models across multiple client engagements.
• Owned the full process from data ingestion and feature mapping through model training, evaluation, and deployment.
• Automated recurring data preprocessing steps, reducing onboarding time across client engagements.