Generative AI Specialization
Five months live from model foundations to shipping GenAI systems: transformers and LLMs, RAG design, fine-tuning (LoRA/QLoRA), multimodal workflows, agents, quantization, and serving. Deeper than Gen AI for Developers. Not a Power BI class.
Compare our Generative AI course in Gurgaon, including both live tracks, published fees, and Sector 68 counselling.
Program fee
One-time payment
The fee is ₹64,999
fee is ₹64,999
What you learn
Who it is for
Career support
Course Curriculum
Module 1: LLM architectures
Model foundations
Specialization starts under the API. You need enough transformer literacy to choose context, tokenize honestly, and know why a model refuses a long doc.
Transformers and attention
- Encoder/decoder vs decoder-only in practice
- Attention as routing, not magic
- A tiny from-scratch or toy pass so the words mean something
Context windows
- What actually fits; what gets dropped
- RoPE-class length tricks at a conceptual level
- When to RAG instead of stuffing the prompt
Tokenization
- Subwords, multilingual pain, code vs prose
- Count tokens before you price a feature
- Truncation bugs that look like ‘the model is dumb’
Behaviour
- Temperature, top-p, and when greedy is correct
- Hallucination as a decoding + data problem
- This is still not a Power BI class
You leave with a one-pager: architecture family, context limit, tokenizer gotchas for your use case.
Module 2: Multimodal systems
Model foundations
Text-only is not the whole job. You will wire vision-language pieces into a workflow you can describe, not a research bake-off.
Vision-language models
- Image in, text out — and the reverse where it exists
- Resolution, tiling, and cost
- When OCR + LLM is more honest than a VLM
ViTs and VLMs at a glance
- Patch embeddings as an idea
- Frozen vs fine-tuned vision towers
- Failure: charts, handwriting, tiny text
Multimodal workflows
- Doc → images → retrieval → answer
- Audio/transcript as just another document
- Eval: did it read the figure or bluff?
Scope control
- You will not train a new CLIP from scratch here
- You will ship a pipeline with tests on 10 messy files
- If you only wanted chat APIs, the developer slab is shorter
You leave with a multimodal pipeline on a small messy corpus and a list of files it fails on.
Module 3: RAG design
Applied GenAI
RAG here is design patterns and evaluation, not a single vendor tutorial. You will choose chunking, retrieval, and grounding like an engineer.
Design patterns
- Naive RAG vs hierarchical vs agentic retrieve
- Hybrid search when embeddings miss keywords
- Query rewrite without infinite loops
Retrieval evaluation
- Recall@k on a labelled set
- Citation rate vs correctness
- Index freshness
Grounding
- Force quotes; fail closed
- Table and PDF layout hell
- Access control as a retrieval filter
When not to RAG
- The answer is in a database — use SQL
- The policy is three paragraphs — put it in the system prompt
- Fine-tune later if the style is the product
You leave with a RAG design doc plus scores on a golden set — not a screenshot of a chat.
Module 4: Fine-tuning
Applied GenAI
PEFT when prompting is not enough. You will run a LoRA/QLoRA-class job, measure it, and know when to throw it away.
Fine-tuning and PEFT
- Data you actually have vs data you wish you had
- LoRA rank, epochs, and overfitting the style
- QLoRA when VRAM is the constraint
Quantization
- Why 4-bit exists
- Quality vs speed vs memory
- Eval after quant — not only before
Preference alignment
- DPO-class idea at a practical level
- Reward hacks you can cause with bad prefs
- Safety evals that are not optional
Decision rule
- Prompt → RAG → fine-tune, in that order unless you have a reason
- A write-up: what improved, what got worse
- Not a dashboard course
You leave with a PEFT run, before/after eval, and a recommendation to keep or drop the adapter.
Module 5: Agentic deployment
Deployment
Serving, cost, and agents that are allowed in production. Deeper than the three-month developer slab; still not LLMOps as a full ops career.
Serving and monitoring
- API vs self-host tradeoffs
- Latency budgets and batching
- A health metric that is not ‘the demo worked’
Cost controls
- Max tokens, caching, cheaper models for easy traffic
- Kill switches
- A weekly cost number in the README
Agentic deployment
- Tool allowlists in prod
- Human-in-the-loop on money/PII
- Rollback a prompt or adapter like a binary
Handoff
- What LLMOps owns next (eval gates, traces, token budgets)
- What you still own (the model choice and the task)
You leave with a served endpoint (or a sober local serve), cost notes, and an agent policy page.
Frequently asked questions
Is this a dashboard or Power BI course?
No. Power BI is only on Data Analytics with AI.
What is the fee and duration?
₹64,999, one-time, 5 months live.
Should I take Gen AI for Developers instead?
If you mainly want to call APIs and ship a RAG UI, take the 3-month developer slab. Take this if you want model and system depth.
Ready to start?
Talk to Prerna about this program — 15 minutes, no sales pitch.