Keerti Purswani - Agentic AI System Design

Keerti Purswani - Agentic AI System Design

Description:

- Core Objective: To teach learners how to design, build, and scale production-grade AI systems utilizing Retrieval-Augmented Generation (RAG), autonomous agents, memory management, evaluation frameworks, and LLMOps.

- Core Philosophy: Shifts the focus from basic prompt engineering to systemic thinking—designing robust agents, workflows, and architectures for real-world, enterprise-level applications.

2. Target Audience

- Software Development Engineers, AI Practitioners, and System Architects who want to move beyond theoretical buzzwords and gain hands-on expertise in building, architecting, and deploying scalable agentic AI systems.

3. Detailed Curriculum Breakdown

The course is structured into 5 comprehensive, progressive modules:

Module 1: Foundation + Knowledge Systems

- Retrieval fundamentals: embeddings, chunking, and vector search.

- Lexical, semantic, and hybrid retrieval techniques.

- Vector databases and nearest-neighbor algorithms (ANN, FAISS, HNSW).

- Production RAG architecture and advanced RAG patterns.

- Metadata filtering, access control, and multi-tenant retrieval.

- RAG debugging, optimization, and evaluation (RAG Evals).

Module 2: Context Engineering

- Context architecture and dynamic context assembly strategies.

- Implementation of short-term and long-term memory (including semantic and user preference memory).

- Token budgeting and context optimization.

- Context compression, summarization, and selective context loading.

- Strategies for managing stale or conflicting context.

Module 3: Agent Systems & Harness Engineering

- Agent architectures and execution patterns.

- Harness engineering: building skills and secure execution environments.

- Tool calling, function orchestration, and MCP (Model Context Protocol) interoperability.

- Tool reliability and fallback handling mechanisms.

- Multi-agent architectures, human-in-the-loop approvals, and state persistence.

Module 4: Production AI Engineering / LLMOps

- Managing long-running agentic workflows.

- Observability, tracing, and debugging in production environments.

- Building automated evaluation pipelines.

- LLMOps and release management workflows.

- Cost optimization, token economics, prompt caching, and semantic caching.

- Model routing, model ensembles, system reliability, security, and guardrails.

Module 5: End-to-End System Design Capstones

- Hands-on design and implementation of an AI coding agent.

- Production architecture discussions applied to real-world AI applications.

4. Key Takeaways & Practical Outcomes

- Systemic Mental Model: Learn to approach AI problems by designing complete systems, agents, and workflows rather than just focusing on isolated prompts.

- Hands-on Project Experience: Build complex, multi-agent applications. A highlighted example is building a SaaS Infrastructure Agent (a multi-agent CLI) that can:

* Design infrastructure directly from natural language requirements.

* Generate and validate Terraform and Docker artifacts.

* Monitor the generated infrastructure and publish artifacts to GitHub.

* Support long-term memory, human approvals, and resumable workflows.

- Production Readiness: Gain practical expertise in making AI systems reliable, observable, cost-effective, and secure for real-world deployment.

View File