Ashton Mehmood

Led end-to-end architecture and implementation of production AI/ML platforms. Built production RAG pipelines (dense vector search, hybrid retrieval, reranking) to improve relevance and grounding. Designed agentic workflows using LangGraph/LangChain with structured tool calling to connect AI with system actions. Oversees MLOps (CI/CD, reproducible training, artifact registries) and deployed low-latency LLM serving (quantization, vLLM). Implemented monitoring for drift and latency/semantic degradation, established automated evaluation suites for RAG quality, fine-tuned LLMs with PEFT/LoRA/QLoRA, and added safety/governance guardrails. Mentored 10–12 engineers and partnered with stakeholders to deliver AI roadmaps.

Owned ML solutions end to end from EDA and feature engineering to deployment. Built and optimized predictive and deep learning models using PyTorch, TensorFlow, and scikit-learn. Containerized and deployed inference microservices using FastAPI, Docker, and Kubernetes across AWS and Azure. Established experiment tracking and model registry workflows with MLflow. Implemented validation and monitoring for anomalies, feature skew, and model degradation. Improved latency and compute efficiency via Optuna tuning, pruning, and ONNX optimization. Engineered distributed feature/data pipelines with PySpark, explored transformer/LLM applications for NLP tasks, and added automated regression/integration tests with pytest. Mentored 6–7 engineers and documented deployment procedures and limitations.

Built data-processing pipelines and exploratory analysis workflows using Python, Pandas, NumPy, and SQL. Developed supervised and unsupervised ML models for classification/regression. Cleaned, transformed, normalized, and validated multi-source structured data for training pipelines. Implemented NLP and foundational deep learning models for text classification using TensorFlow, PyTorch, and NLTK. Deployed trained models as reusable REST APIs with Flask/FastAPI. Evaluated with precision, recall, F1, ROC-AUC, and cross-validation; performed hyperparameter tuning for model selection. Maintained production-ready code and version-controlled experiments with Git.