Amador Jr B B

Fine-tuned open-source LLMs (LLaMA, Mistral) using HuggingFace and LangChain, improving clinical decision-support accuracy by ~12%. Built retrieval-augmented generation (RAG) systems with Azure AI Search and Pinecone, achieving 25% higher precision on clinical knowledge queries. Developed multimodal GenAI prototypes combining text, speech, and emotion signals for clinician-facing tools. Benchmarked proprietary LLMs against fine-tuned open-source models, contributing to cost/performance analyses influencing $1M+ annual vendor decisions. Deployed low-latency inference pipelines across Azure AI, AWS SageMaker, and GCP Vertex AI, reducing response time by ~40%. Designed and maintained FastAPI and gRPC services to integrate AI recommendations into clinical systems serving 10K+ daily users. Optimized distributed training using Ray and GPU acceleration, cutting training time by ~30%. Implemented MLOps workflows (CI/CD, MLflow, monitoring) to ensure reproducibility, observability, and complian

Built predictive models (Random Forest, XGBoost, CNNs, RNNs) for patient risk scoring and readmission prediction, reducing error rates by ~20%. Developed deep learning pipelines for MRI and CT scan analysis, achieving ~92% diagnostic accuracy in pilot studies. Migrated NLP pipelines from Word2Vec/GloVe to BERT-based models, improving clinical text classification accuracy by ~18%. Designed large-scale ETL pipelines with Spark and Airflow processing terabytes of healthcare data daily. Applied synthetic data generation to improve model performance on rare medical conditions. Delivered explainability dashboards (SHAP, LIME) to clinicians, improving trust and adoption of ML outputs. Exposed ML models via REST APIs and integrated them into hospital IT systems. Collaborated with clinicians, data engineers, and compliance teams to ensure medical relevance and regulatory alignment.

Developed fraud detection pipelines with Python, Scikit-learn, and TensorFlow, reducing false positives by 28% and increasing approval rates by 15%. Built real-time anomaly detection systems using Kafka and Python, scaling to 50K+ daily transactions. Implemented NLP sentiment analysis models with Word2Vec and GloVe, later migrated to HuggingFace Transformers, boosting accuracy by 15%. Containerized ML services with Docker and Kubernetes (early adoption), achieving 99.9% uptime for secure financial APIs. Designed CI/CD pipelines with Jenkins and GitLab CI, reducing deployment failures by 40%. Integrated ML outputs into compliance workflows, improving audit readiness. Deployed monitoring with Prometheus and Grafana, enabling real-time tracking of cost, latency, and pipeline reliability. Contributed reusable ML libraries and utilities.