Loading demo links...
What to study for each NCA-AMAD domain
Agentic AI
Covers AI agent concepts including planning, task decomposition, tool and function calling, memory, orchestration, and multi-agent systems. It also addresses retrieval-grounded agents, reflection, human-in-the-loop control, guardrails, and evaluating agent task success.
Key skills
Agent planning and task decomposition, tool and function calling, short- and long-term memory, multi-agent orchestration, and guardrails for autonomous behaviour.
Data Processing and Data Flywheel
Covers preparing data for AI, including cleaning, tokenisation, annotation, deduplication, synthetic data, and train, validation, and test splits. It also addresses the data flywheel, active learning, human feedback, PII handling, dataset versioning, and monitoring for data drift.
Key skills
Cleaning, tokenising, annotating and deduplicating training data, synthetic data generation, and the data flywheel loop that improves models from production feedback.
Inference Concepts
Covers running trained models in production, including the latency and throughput trade-off, dynamic batching, the KV cache, and quantisation. It also addresses inference optimisation with tools such as TensorRT and Triton Inference Server, streaming and time to first token, speculative decoding, autoscaling, and edge versus cloud inference.
Key skills
Latency versus throughput trade-offs, dynamic and continuous batching, KV cache, streaming tokens, and inference server basics.
Model Customization and Fine-Tuning Concepts
Covers adapting models through prompt engineering, retrieval-augmented generation, and fine-tuning, including full fine-tuning and parameter-efficient methods such as LoRA. It also addresses instruction tuning, RLHF, overfitting and catastrophic forgetting, learning-rate choices, SFT data formats, and held-out evaluation.
Key skills
Prompt engineering, retrieval-augmented generation, instruction tuning and parameter-efficient fine-tuning such as LoRA, and choosing between them.
Models and Model Types
Covers the main model families used in AI applications, including large language models, vision models, multimodal models, embedding models, and diffusion models. It also addresses foundation models, small language models, encoder and decoder transformers, and speech models for recognition and synthesis.
Key skills
Large language models, vision and multimodal models, embedding and reranking models, speech models, and matching model types to application needs.
NVIDIA AI Enterprise and Blueprints Overview
Covers NVIDIA's AI software platform, including NVIDIA AI Enterprise for production support and validated, optimised software, and NVIDIA NIM inference microservices for API-based deployment. It also addresses NVIDIA AI Blueprints as customisable reference workflows, the NeMo framework for model customisation, and CUDA as the foundation of GPU acceleration.
Key skills
NVIDIA AI Enterprise support and validated software, NIM microservices, and NVIDIA Blueprints as reference workflows for common AI applications.
Sizing and Configuration
Covers sizing accelerated infrastructure for AI workloads, including GPU memory requirements from model size and precision, and capacity planning from request rate, concurrency, and token throughput. It also addresses tensor parallelism, batch size and memory trade-offs, inter-GPU interconnect bandwidth, GPU selection, and benchmarking before production.
Key skills
Estimating GPU memory from model size and precision, selecting GPU counts for training versus inference, and configuring accelerated nodes for AI workloads.
Physical AI
Covers AI that perceives and acts in the physical world through systems such as robots, including simulation-based development and the sim-to-real workflow. It also addresses sensor fusion from cameras, lidar, and other sensors, and digital twins for testing and optimising physical environments.
Key skills
Perception-action loops in robotics, simulation-based training and sim-to-real, sensor data, and NVIDIA Isaac/Omniverse concepts at an introductory level.