Generative AI & LLM Course: live online in India

A live generative AI and large language model course: tokenizers, attention and transformers from scratch, building a GPT-style LLM, fine-tuning with LoRA/QLoRA, RAG and AI agents, then research-level LLM engineering.

  • 16 weeksof generative AI and LLMs inside Program Zero
  • 9 PM ISTlive, Mon–Fri · recorded
  • ₹89,999or ₹7,500/month EMI
  • 9 January 2027next batch starts

Generative AI & LLM Course in Program Zero is taught live by working engineers as part of an 18-month, mentor-led programme. This is where Program Zero goes further than most courses. Over 16 weeks you derive attention, build a transformer from scratch, write your own tokenizer and train a small GPT-style model in PyTorch. Then you fine-tune open models with Hugging Face and LoRA/QLoRA, build retrieval-augmented generation with vector databases, and connect models to tools as agents.

The advanced phase covers what frontier teams work on: large-scale data pipelines, distributed training (FSDP, DeepSpeed ZeRO), scaling laws, modern architectures (RoPE, GQA, mixture of experts), inference optimisation (quantization, vLLM) and alignment (SFT, PPO, DPO, red-teaming).

Who it's for

  • Learners who want to understand and build LLMs, not only call APIs
  • Developers building GenAI features into products
  • Anyone aiming at LLM engineer or AI research engineer roles

What you'll be able to do

  • Build a transformer and a small GPT from scratch
  • Fine-tune open LLMs with LoRA/QLoRA
  • Build RAG pipelines and tool-using agents
  • Optimise inference and understand alignment methods

Syllabus

What the generative AI and LLMs course covers, week by week

Taken directly from the Program Zero curriculum. Weeks are counted from the batch start on 9 January 2027.

Phase 8 · weeks 54-61 · 8 weeks

NLP, Transformers & Building LLMs (Core)

  • NLP fundamentals: tokenization, TF-IDF, embeddings (Word2Vec, GloVe)
  • Attention mechanism derivation; seq2seq
  • Transformer architecture from scratch: self-attention, multi-head attention, positional encoding, encoder-decoder, layer norm, residuals
  • Tokenizers in depth: BPE, WordPiece, SentencePiece — build a custom tokenizer
  • Build a GPT-style decoder-only LLM from scratch in PyTorch: embeddings → transformer blocks → output head, causal masking, training loop, cross-entropy loss, perplexity
  • Train a small LLM end-to-end on a custom dataset
  • BERT / encoder-only models (masked language modeling)
  • Fine-tuning with HuggingFace Transformers: classification, NER, summarization, Q&A
  • LoRA / QLoRA parameter-efficient fine-tuning
  • RAG: embeddings, vector DBs (FAISS/Chroma/Pinecone), building a RAG pipeline
  • Prompt engineering, chain-of-thought, AI agents/tool use (LangChain/LlamaIndex)
  • Evaluating LLMs: perplexity, BLEU/ROUGE, benchmarks

Project: Small GPT trained from scratch on a custom corpus; LoRA fine-tune of an open model (Llama/Mistral/Qwen family)

Phase 9 · weeks 62-69 · 8 weeks

Advanced LLM Engineering (Research Level)

  • Web-scale data collection (Common Crawl-style), deduplication (MinHash + LSH), quality filtering (heuristic + classifier-based), toxicity/PII filtering
  • Data mixing ratios, curriculum ordering; tokenizer training at scale
  • Build a full pipeline: raw data → clean → dedupe → filter → tokenize → shard
  • DDP internals, gradient sync, communication overhead
  • FSDP and DeepSpeed ZeRO (stages 1/2/3) — sharding optimizer states/gradients/parameters
  • Tensor parallelism & pipeline parallelism (Megatron-LM style); 3D parallelism
  • Gradient checkpointing, mixed precision (fp16/bf16) at scale
  • NCCL basics, multi-node orchestration (Slurm/Kubernetes training jobs), checkpointing & fault tolerance for long runs
  • Kaplan et al. and Chinchilla scaling laws
  • Compute-optimal trade-offs: model size vs. data size vs. compute budget
  • Estimating cost/time before committing a training run; designing an ablation plan
  • RoPE, ALiBi (why they replaced sinusoidal encoding)
  • Grouped-Query Attention (GQA), Multi-Query Attention (MQA)
  • Sliding-window/sparse attention, long-context techniques
  • Mixture-of-Experts (MoE): routing, load balancing, sparse activation
  • Reading/implementing components from Llama, Mistral, DeepSeek, Qwen papers
  • KV-cache management; quantization: GPTQ, AWQ, GGUF/llama.cpp, int8/int4
  • Speculative decoding, continuous batching
  • Production serving: vLLM, TensorRT-LLM, TGI
  • Latency/throughput/cost-per-token trade-offs
  • SFT pipeline design and instruction-data curation
  • Reward model training: preference data, Bradley-Terry loss
  • PPO implementation for RLHF (hands-on)
  • DPO and modern alternatives — implementation
  • Red-teaming methodology, safety evaluation, constitutional AI concepts, guardrail design
  • Reading and reproducing a research paper end-to-end
  • Designing and running ablation studies; rigorous experiment comparison
  • Writing a technical report on a training run/ablation
  • Keeping up with the field critically: key labs, papers, benchmarks

Project: Execute a full, faithful (small-scale) pretraining run — build the data pipeline, choose an architecture variant, run compute-optimal sizing analysis, train with a real distributed setup (2–8 GPUs), run an SFT + DPO alignment pass, then serve the model with an optimized inference stack. Written up as a technical report exactly as a research team would produce.

How you're taught
Live classesMonday to Friday, 9:00–10:30 PM IST, every class recorded
PracticeSaturday assignments and live projects, reviewed by mentors
HelpSunday 11:00 AM–1:00 PM doubt-clearing session
AssessmentA cumulative test every 3 months (retest a week later); certificates on merit
Fee₹89,999 one-time, or ₹7,500 × 15 months by EMI (₹1,12,500 total), inclusive of all taxes, covering every subject

Book your seat

How to join in 4 steps

  1. 1 Create your account Sign up with your name, email and Indian mobile number.
  2. 2 Verify your email Enter the 6-digit code we email you.
  3. 3 Pay the membership fee Pay ₹89,999 one-time, or ₹7,500 to start an EMI plan (₹7,500 × 15 months), securely with Razorpay (UPI, cards, net banking).
  4. 4 Join the first live class Classes start on 9 January 2027, Monday to Friday, 9:00–10:30 PM IST.

Book your seat

Questions

Generative AI & LLM Course: questions

Will I really build an LLM from scratch?

Yes, a small GPT-style model trained end to end on a custom dataset. It is small enough to train on modest hardware and teaches exactly how the large ones work.

Does the generative AI course cover RAG and agents?

Yes: embeddings, vector databases (FAISS, Chroma, Pinecone), building a RAG pipeline, prompt engineering, and tool-using agents with LangChain or LlamaIndex.

Is prompt engineering enough for a GenAI career?

No. Prompting is one skill. Building GenAI products also needs programming, retrieval, fine-tuning, evaluation and deployment, which this course covers.

How much does the generative AI and LLMs course cost?

Generative AI & LLM Course is part of Program Zero: ₹89,999 one-time for all 18 months, or ₹7,500 × 15 months by EMI (₹1,12,500 in total), inclusive of all taxes. Every other subject is included.

When does the next generative AI and LLMs batch start?

The next batch starts on 9 January 2027. Live classes run Monday to Friday, 9:00–10:30 PM IST.

Learn generative AI and LLMs live, from 9 January 2027.

₹89,999 one-time or ₹7,500/month · seats are limited.

Book Your Seat

Install Program Zero on iPhone or iPad

  1. Open zero.careersninza.com in Safari.
  2. Tap the Share button .
  3. Scroll down and tap Add to Home Screen, then Add.

The app opens full-screen from your home screen, just like an App Store app.

Install Program Zero on Android

  1. Open zero.careersninza.com in Chrome.
  2. Tap the ⋮ menu at the top right.
  3. Tap Install app (or Add to Home screen), then Install.

You're already using the app

Program Zero is installed on this device.