Srinadh Y

Fine-tuned Llama 3 models for large-scale content policy enforcement, improving harmful content detection by 25% and reducing false positives by 15% using LoRA, PEFT, Hugging Face, PyTorch Lightning, ONNX Runtime, and W&B. Designed and deployed autonomous, tool-using AI agents using LangChain, FAISS, and Google ADK, enabling real-time reasoning, retrieval, and workflow automation — reducing manual interventions by 40%. Built retrieval-augmented generation (RAG) pipelines over large document stores and APIs to improve moderation accuracy and decision latency at scale. Scaled training and inference on Meta Research SuperCluster (RSC) using FSDP, FlashAttention, MLX, and Data Swarm, cutting infrastructure costs by 30% and reducing latency. Developed diffusion-based generative models for dynamic ad creative generation, driving +12% user engagement and +8% CTR across Meta ad surfaces. Implemented bias detection and fairness evaluation pipelines using SHAP, Fairness Indicators, and A/B testing, reducing demographic performance gaps by 22% without impacting engagement. Engineered Spark/PySpark ETL pipelines processing 5TB+ daily data, including anomaly detection (Isolation Forest) to ensure training data quality and improve model accuracy by 9%. Collaborated closely with Trust & Safety, Policy, Product, and AI Infrastructure teams to align AI systems with evolving content standards and product goals.

Built predictive ML models in Python and R to forecast device failure risk and customer churn, improving model accuracy by 18% and reducing false negatives by 22%. Developed and deployed LLM-powered internal AI agents using LangChain, Google Gemini, and Vertex AI, automating HR and IT support workflows and reducing ticket response time by 40%. Designed RAG-based, time-sensitive agent systems combining operational metrics forecasting (ARIMA, Prophet) with real-time alerts for demand and capacity planning. Implemented ethical AI and fairness checks using SHAP and experimentation frameworks, ensuring bias-aware decision-making across automated workflows. Conducted model performance monitoring and drift analysis using statistical testing and explainability tools, improving model stability by 15%. Evaluated and optimized models using AUC, F1, RMSE, and GridSearchCV, ensuring strong generalization across Apple’s diverse product lines. Partnered with data engineering teams to build CI/CD pipelines and integrate ML solutions with SQL and NoSQL platforms for scalable production deployment.

Designed and deployed ML models (classification, regression, NLP) using Python, TensorFlow, and Scikit-learn, improving prediction accuracy by 15–20% across enterprise use cases. Built a personalized recommendation engine using collaborative filtering and deep learning, increasing customer engagement by 12%. Developed end-to-end ML pipelines on AWS SageMaker, integrating Redshift and Athena for feature engineering and analytics, reducing training time by 30%. Implemented CI/CD pipelines with Git, Jenkins, and GitHub Actions to automate model testing and deployment, cutting release cycles by 40%. Built NLP-based chatbots using BERT and Transformer architectures, reducing customer query resolution time by 40% for BFSI clients. Created Tableau dashboards to monitor ML performance and operational KPIs, identifying optimization opportunities that reduced client costs by 18% annually.