AI/ML Engineer at AMD

Built RAG applications for engineering knowledge search over 200K+ internal documents using LangChain/LlamaIndex, FAISS, and Pinecone to provide grounded answers. Developed LangGraph-based AI agents with MCP integrations to securely perform multi-step lookups across internal tools and documentation services. Created training data pipelines with Airflow and Spark for cleaning, deduplication, and chunking, keeping the retrieval index refreshed nightly. Fine-tuned open-weight LLMs (e.g., Llama) with LoRA/PEFT using distributed multi-GPU training (PyTorch FSDP/DeepSpeed ZeRO), improving domain accuracy from 72% to 88% and reducing training time by ~30% on 7B models. Tuned retrieval (chunking/reranking/hybrid search) across 25+ configurations to reduce off-topic/unanswered responses by ~1/3. Implemented LLM evaluation checks for retrieval relevance, faithfulness, and latency across 200 benchmark questions. Deployed models on Kubernetes with KServe/FastAPI, MLflow, GitHub Actions, Redis cach

Machine Learning Engineer at Cognizant

Built and shipped ML services for banking and retail clients using PyTorch, scikit-learn, and XGBoost, improving accuracy by ~15% validated via A/B testing (F1, AUC-ROC) and SHAP-based governance. Fine-tuned BERT and Hugging Face Transformers for document/ticket classification to reach 93% accuracy using mixed precision and multi-GPU data parallelism. Prototyped a document QA service using OpenAI embeddings and vector search over 5K+ policy documents, reducing analyst time spent searching PDFs by ~40%. Trained computer vision models (PyTorch, transfer learning, ResNet) to auto-tag 12K+ SKUs in a product catalog and helped migrate/maintain inherited TensorFlow/Keras models. Rebuilt Spark/AWS Glue/Airflow pipelines handling 10M+ records/day into Snowflake/MongoDB; runtime reduced by ~35% via skew handling and feature caching. Packaged inference services across 4 client environments on multiple clouds using Docker/Kubernetes/Azure ML/Vertex AI and tracked experiments with MLflow across 30