Post Graduate Certificate in Agentic and RAG Systems for AI Engineering

• Module 2: LLM Architecture and Model Ecosystem

  • Transformer architecture, attention mechanism, tokenization, embeddings, positional encoding

  • Context windows, GPT/Llama/Mistral ecosystem

  • Open-source vs proprietary models, structured outputs, streaming responses

  • Cost-latency-quality tradeoffs

• Module 3: Prompt Engineering, Evaluation and Reliability

  • Zero-shot and few-shot prompting, chain-of-thought reasoning, tool-aware prompting

  • Retrieval-aware prompting, system prompts, safety prompts, hallucination mitigation

  • Prompt templates, prompt versioning

  • LLM evaluation frameworks; LLM-as-judge, pairwise comparison, critic-creator loop

  • Golden datasets, RAGAS/DeepEval eval stack, reliability testing

• Module 4: Retrieval Architecture and Vector Infrastructure

  • RAG fundamentals, embedding models: OpenAI/BGE/Nomic

  • Semantic similarity, chunking strategies

  • Vector databases: Qdrant/FAISS/Chroma, HNSW and IVF indexing

  • Metadata enrichment, ingestion pipelines, hybrid retrieval systems

  • Knowledge lifecycle management, PII detection and redaction

• Module 5: Production RAG and Retrieval Optimisation

  • BM25 and dense retrieval, reranking model: cross-encoder

  • Query rewriting, HyDE retrieval, multi-query retrieval

  • Semantic caching, context compression

  • Groundedness evaluation: RAGAS/DeepEval/TruLens, precision@K

  • RAG failure taxonomy: retrieval vs generation vs prompt, retrieval debugging

• Module 6: Fine-Tuning Strategy and Adaptation Frameworks

  • Prompting vs. RAG vs. fine-tuning: choosing the right approach for different use cases

  • Dataset curation and instruction tuning for domain-specific AI applications

  • Domain adaptation and synthetic data generation for model improvement

  • Baseline evaluation, model selection, and tuned vs. base model comparison

  • Cost, performance, and risk analysis for AI model customization

  • Identifying when fine-tuning is the wrong choice and selecting better alternatives

• Module 7: Parameter-Efficient Model Adaptation

  • LoRA, QLoRA, PEFT frameworks, quantization, adapter architectures

  • Training workflows, hyperparameter tuning, dataset formatting

  • Model evaluation, serving fine-tuned models

  • HuggingFace PEFT/TRL, bitsandbytes, Unsloth/Axolotl

  • Catastrophic forgetting, GPU/Colab lab path

• Module 8: Agent Architecture, Tooling and Memory Systems

  • ReAct architecture, tool-calling agents, API orchestration, function calling

  • Planning loops, memory systems, vector memory, retrieval-as-a-tool

  • Grounded agents, text-to-SQL agents

  • Failure-Oriented Design: transient/permanent/silent/cascading/adversarial failures

  • Max-iteration guards, circuit breakers, prompt-injection defense

• Module 9: Agent Orchestration and Human Oversight

  • LangGraph workflows, state-machine orchestration, planner-executor systems

  • Reflection loops, self-correction systems

  • HITL workflows, approval gates, escalation systems, audit trails

  • Streaming UX; tool-call streaming, reasoning streaming, interruptibility

• Module 10: Multi-Agent Coordination Systems

  • Multi-agent architectures, supervisor-worker systems, task delegation

  • Shared state management, inter-agent communication, coordination failures

  • Distributed workflows, cost and latency optimisation

  • When NOT to use multi-agent systems

• Module 11: MCP Protocols, AI Security and Governance

  • MCP architecture, SSE and stdio transport

  • Retrieval poisoning, tool abuse vulnerabilities, data exfiltration risks

  • AI guardrails: Guardrails AI, NeMo Guardrails, permission management

  • GDPR and DPDP compliance, enterprise AI governance, audit trails

• Module 12: AI Deployment, Observability and Runtime Operations

  • Logging, tracing and metrics; Langfuse

  • Distributed tracing, cost observability, token tracking

  • Semantic caching for cost, FastAPI deployment

  • CI/CD for AI systems; GitHub Actions, regression testing

  • Prompt versioning, model routing; runtime fallback mechanisms

• Module 13: Capstone Project

  • Architecture design, RAG integration, agent workflow implementation

  • Fine-tuned model integration, deployment pipeline setup

Week 9 | Enterprise RAG and Retrieval Infrastructure

You build your first end-to-end RAG pipeline from scratch over a corpus you choose. Load, chunk, embed, store, retrieve, and generate answers grounded in what you put in. The first time it actually works on a document you uploaded yourself is a different feeling from a tutorial.

You will build: A working end-to-end RAG pipeline from document ingestion to grounded answer generation.

Week 12 | Production RAG and Retrieval Optimisation

You take Mini-RAG further add hybrid retrieval, a reranker, inline source citations, and confident abstention when the answer isn't supported. Then you measure it: hit rate and precision@K on a real question set. This is where RAG stops being a demo.

You will build: A production-ready document assistant with citations, abstention, hybrid retrieval, and a retrieval quality report.

Week 14 | Model Adaptation and Fine-Tuning

You take a small open-source model — Llama, Mistral, or Phi-class, curate a dataset for one narrow task, and run a QLoRA fine-tune on a GPU. Then you evaluate the tuned model against the base model and write a decision log: when was fine-tuning the right call and when wasn't it.

You will build: A fine-tuned open-source model with evaluation results and a written decision log.

Week 20 | Agentic AI and Autonomous Workflows

You build a single agent that plans and uses two or three real tools — document search, calculator, SQL lookup. You implement a visible reasoning trace with max-iteration guards and failure fallback. The optional extension: a small planner-worker-reviewer crew with clear roles and handoffs.

You will build: A working tool-calling agent with a visible reasoning trace, iteration guards, and failure fallback — deployed and callable.

Week 21 | AI Infrastructure, Security and Governance

You connect an assistant to an external data source through an MCP server building or connecting a small custom server and securing it with an allow-list or input validation. This is the architecture pattern behind every serious enterprise AI integration happening right now.

You will build: A live MCP-connected assistant with a secured, exposed tool; callable, demonstrable, and ready to extend.

The Certificate of Completion is issued by iHub DivyaSampark, IIT Roorkee, a Government of India initiative under the Department of Science and Technology. It carries the institutional credibility of IIT Roorkee, one of India's oldest and most reputed technical universities, and the government backing of the NM-ICPS mission. For technology employers evaluating candidates in AI and software engineering, the combination of the IIT Roorkee name and the practical, production-focused curriculum rather than theoretical coursework makes the credential meaningful. The certificate validates demonstrated skills across RAG systems, agentic AI, fine-tuning, and deployment, not just course attendance.

To be eligible for the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering, applicants should hold a bachelor's degree or higher. Diploma holders with a minimum of three years of professional work experience are also eligible to apply. In addition, intermediate familiarity with programming concepts is expected through prior coding experience.

The Post Graduate Certificate in Agentic and RAG Systems for AI Engineering is designed for two primary audiences. The first is software engineers, data professionals, and AI practitioners who build software professionally, have worked with APIs and codebases, and want to develop the ability to build production-grade AI systems not just use AI tools. The second is tech leads, architects, and product managers who make technology decisions, have prior exposure to software or AI projects, and want the technical depth to evaluate AI proposals, guide implementation, and drive adoption confidently. Both audiences benefit from the programme's practical, engineering-first structure.

Participants in the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering should expect to spend approximately 8 to 10 hours per week on programme activities. This includes attendance at two live sessions per week, each three hours long, plus time for assignments, hands-on labs, and project work. The programme is specifically designed to be compatible with a full-time working schedule. No career break is needed to complete this programme; it has been built for professionals who are learning while continuing to work.

Participants joining the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering are expected to have an intermediate understanding of programming concepts. The programme begins with a dedicated Python for AI Engineering module to bring all participants to a consistent baseline in production-oriented Python covering async programming, API handling, Pydantic validation, and pipeline-oriented coding. The programme is not designed for complete beginners to programming but does not require prior deep software engineering expertise.

During the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering, participants will complete six module-end projects, each tied to a specific pillar of the curriculum, plus a final capstone project. The module projects cover prompt evaluation, RAG pipeline construction, production RAG optimisation with hybrid retrieval and reranking, LLM fine-tuning with QLoRA, tool-calling agent development, and MCP-based AI integration with security guardrails. The capstone requires participants to build, deploy, and monitor a full production AI system integrating RAG, an autonomous agent, a fine-tuned model, and a observability dashboard. All projects are designed to mirror real engineering tasks, not tutorial exercises.

Upon completing the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering, participants who meet all completion requirements will receive a Certificate of Completion from iHUB DivyaSampark, IIT Roorkee. The certificate is issued by iHub DivyaSampark, a Government of India initiative, and carries the IIT Roorkee institutional identity. To receive the certificate, participants must complete all graded mini projects with a minimum overall score of 50%, achieve at least 50% in the capstone project, and maintain a minimum of 50% attendance throughout the programme. The certificate is awarded as an e-certificate upon successful completion.

The Post Graduate Certificate in Agentic and RAG Systems for AI Engineering provides hands-on exposure to 15+ industry-standard tools across the full AI engineering lifecycle. For RAG and retrieval, participants work with Qdrant, FAISS, Chroma, and OpenAI embeddings. For agentic AI, the programme covers LangGraph, CrewAI, and Smolagents. For fine-tuning, participants use Hugging Face, Unsloth, and Axolotl. Deployment and observability tools include FastAPI, GitHub Actions and Langfuse. Security and governance modules cover Guardrails AI and NeMo Guardrails. The programme also includes Model Context Protocol (MCP) integration, a skill most comparable programmes do not cover.

The capstone project in the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering requires participants to design, build, and deploy a production-ready AI application that integrates the core technical capabilities covered throughout the programme. Specifically, participants build a production-grade RAG pipeline with hybrid search and reranking, a LangGraph-orchestrated agent with tool-calling and failure handling, a FastAPI deployment for scalable API serving, and a Langfuse monitoring dashboard tracking latency, cost, and answer quality. The final deliverables include a GitHub repository with the complete system, a deployed API endpoint, a real-time monitoring dashboard, and a demo-ready walkthrough.

Yes. Participants enrolled in the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering have access to an optional two-day on-campus immersion experience at IIT Roorkee's Noida campus. This optional experience is designed to provide a sense of connection to the IIT Roorkee ecosystem beyond the online programme environment.

For the Post Graduate Certificate in Agentic and RAG Systems for AI Engineering, refund requests may be initiated before the commencement of the programme. Once the programme begins on 23 September 2026, the programme fee becomes non-refundable. Participants are advised to review the full terms and conditions before making a payment.