Generative AI Curriculum Index
A rigorous, undergraduate-to-industry engineering curriculum tailored for B.Tech Computer Science, Data Science, and Artificial Intelligence students.
Designed according to university academic standards (AICTE/UGC) and modern industry generative modeling practices,
this program bridges theoretical mathematical foundations (linear algebra, probability, information theory) with hands-on transformer implementation,
large language model (LLM) fine-tuning, Retrieval-Augmented Generation (RAG), diffusion models, multimodal systems, and production LLMOps pipelines.
- Master mathematical prerequisites: high-dimensional linear algebra, multivariate calculus, Bayes theorem, entropy, and KL divergence.
- Understand the evolution from statistical language models and recurrent architectures (RNNs, LSTMs, GRUs) to modern self-attention.
- Implement Scaled Dot-Product Attention, Multi-Head Attention, and full Encoder-Decoder Transformer architectures from scratch in PyTorch.
- Explore pre-training objectives: Masked Language Modeling (BERT), Autoregressive Causal Modeling (GPT), and Sequence-to-Sequence (T5).
- Analyze modern LLM scaling laws, compute-optimal training regimes (Chinchilla), and tokenization algorithms (BPE, WordPiece, SentencePiece).
- Fine-tune foundation models using Parameter-Efficient Fine-Tuning (PEFT) methods: LoRA, QLoRA, Prefix Tuning, and Prompt Tuning.
- Align large models with human preferences using Reinforcement Learning from Human Feedback (RLHF), DPO, PPO, and KTO algorithms.
- Architect enterprise-grade Retrieval-Augmented Generation (RAG) pipelines using vector embeddings, chunking, and hybrid search.
- Deploy and query production vector databases (Pinecone, Milvus, Qdrant, ChromaDB, pgvector) with dense and sparse re-ranking.
- Build autonomous multi-agent cognitive architectures using LangChain, LlamaIndex, LangGraph, and AutoGen.
- Understand probabilistic deep generative models: Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Normalizing Flows.
- Formulate continuous and discrete Denoising Diffusion Probabilistic Models (DDPM, DDIM) and Latent Diffusion Models (Stable Diffusion).
- Engineer multimodal systems combining vision-language models (CLIP, LLaVA, Flamingo) and audio diffusion architectures (Whisper, AudioCraft).
- Optimize inference and production serving using Quantization (AWQ, GPTQ, GGUF), FlashAttention-2, vLLM, and TensorRT-LLM.
- Enforce GenAI safety, guardrails (NeMo Guardrails, Llama Guard), bias mitigation, red-teaming, and intellectual property compliance.