Software engineers
Go beyond calling LLM APIs: run, scale and secure your own inference, retrieval and agent infrastructure.
AI ENGINEER ADVANCED
From production AI
to AI at scale.
A 32-week production syllabus for experienced software and data engineers: self-hosted inference, enterprise retrieval, GraphRAG, stateful agents, fine-tuning and LLMOps, built to deploy robust, cost-effective and observable AI systems at scale.
LOGICRAFT.AIAI ENGINEER
ADVANCEDLEARN · BUILD · SHIP
01 GPUs02 RAG03 Agents04 LLMOpsADVANCEDLEARN · BUILD · SHIP
THE CURRICULUM
6 modules.
32 weeks of production engineering.
Download Syllabus
(The topics in the above syllabus would increase as per industry need/standard)
01 / INFERENCE
Low-Level LLM Systems, Distributed Inference & GPU Kernels
Demystifying GPU memory hierarchies, serving engines, token throughput, and high-scale inference.
02 / RETRIEVALAdvanced Enterprise Retrieval & Vector Architectures
Scaling beyond basic vector search to deterministic, multi-tenant enterprise knowledge retrieval.
03 / AGENTSStateful Agentic Systems & Model Context Protocol (MCP)
Complex multi-agent coordination, deterministic execution graphs, and tool protocols.
04 / FINE-TUNINGDomain Adaptation, Supervised Fine-Tuning & Alignment
When prompting and RAG fall short: custom model training, PEFT, and preference alignment.
05 / LLMOPSLLMOps, Continuous Evaluation, Security & Observability
Industrializing AI deployments with enterprise security, telemetry, and automated CI/CD.
06 / CAPSTONECapstone Architecture & Defense
End-to-end delivery of an enterprise-grade AI system with strict SLA and cost boundaries.
THE CURRICULUM
Week by week.
Class by class.
Six modules, 32 weeks and 96 hands-on classes. Modules 01–04 each end with a production milestone project and architectural review; the program closes with a four-week capstone and architectural defense.
6 modules32 weeks96 classes
Download Syllabus
(The topics in the above syllabus would increase as per industry need/standard)
Module 01 · Weeks 01–06
Low-Level LLM Systems, Distributed Inference & GPU Kernels
Focus: Demystifying GPU memory hierarchies, serving engines, token throughput, and high-scale inference.
01Modern LLM Architecture & Computational Hardware ConstraintsWeek 01 · Classes 01–03+
- Class 01 · Hardware Architectures Tensor Cores, HBM3/HBM3e bandwidth, NVLink topologies, memory-bandwidth bound vs. compute-bound operations.
- Class 02 · FlashAttention Deep Dive IO awareness, GPU SRAM tiling, and avoiding HBM roundtrips in FlashAttention-2 and FlashAttention-3.
- Class 03 · Modern Attention Topologies MHA (Multi-Head), MQA (Multi-Query), and GQA (Grouped-Query Attention) compute and memory trade-offs.
02KV Cache Mechanics, PagedAttention & Continuous BatchingWeek 02 · Classes 04–06+
- Class 04 · The KV Cache Bottleneck Memory footprint formulas, dynamic context allocation, and generation bottlenecks.
- Class 05 · PagedAttention & vLLM Internals Virtual memory paging, block table managers, logical vs physical memory, and prefix caching.
- Class 06 · Dynamic Batching Mechanics Static vs continuous batching, prefill vs decode phase scheduling, request queues, and preemption.
03Distributed Inference Engines (vLLM & TensorRT-LLM)Week 03 · Classes 07–09+
- Class 07 · Production vLLM Configuration Engine flags, max_model_len, GPU memory utilization thresholds, swap space, and chunked prefill.
- Class 08 · TensorRT-LLM Architecture Engine compilation, kernel fusion, weight-only quantization, and low-latency pipeline optimization.
- Class 09 · Distributed Parallelism Strategies Tensor Parallelism (TP) across GPUs vs Pipeline Parallelism (PP) and Megatron-LM communication primitives.
04Model Quantization & Speculative DecodingWeek 04 · Classes 10–12+
- Class 10 · Post-Training Quantization (PTQ) Math and runtime trade-offs of FP8, INT8, INT4, AWQ (Activation-aware Weight Quantization), and GPTQ.
- Class 11 · Speculative Decoding Systems Draft-verification model pipelines, medusa heads, acceptance rates, and token throughput gains.
- Class 12 · Structured Output Acceleration Constrained decoding, Outlines, Guidance, and grammar-guided token masking at the logit level.
05Cloud & Kubernetes-Based GPU AutoscalingWeek 05 · Classes 13–15+
- Class 13 · Containerizing LLM Runtimes NVIDIA Container Toolkit, multi-stage production Dockerfiles, and cold-start reduction techniques.
- Class 14 · Orchestrating GPU Workloads on K8s EKS/GKE device plugins, GPU time-slicing vs MIG (Multi-Instance GPU), and Karpenter provisioning.
- Class 15 · Metric-Based Dynamic Autoscaling Autoscaling on TTFT (Time to First Token), TPOT (Time Per Output Token), and request queue concurrency.
06Production Milestone Project 1Week 06 · Classes 16–18+
- Class 16 · Production Lab 1 Multi-GPU vLLM cluster deployment on Kubernetes with prefix caching and continuous batching enabled.
- Class 17 · Production Lab 2 Load testing using Locust and Triton perf-analyzer; benchmarking latency vs. throughput curves under concurrency.
- Class 18 · Architectural Review Code review, architectural defense, failure-recovery drills, and milestone sign-off.
Module 02 · Weeks 07–12
Advanced Enterprise Retrieval & Vector Architectures
Focus: Scaling beyond basic vector search to deterministic, multi-tenant enterprise knowledge retrieval.
07Deterministic Ingestion & Semantic ChunkingWeek 07 · Classes 19–21+
- Class 19 · Ingestion Failure Modes Pitfalls of fixed-size chunking, boundary fragmentation, and document layout destruction.
- Class 20 · Semantic & Structure-Aware Chunking Embedding distance splitting, sliding windows, and Markdown/HTML hierarchy preservation.
- Class 21 · Agentic & Contextual Chunking Contextual chunk headers, pre-pending LLM-generated document/section summaries, and metadata injection.
08Hybrid Search & Reciprocal Rank Fusion (RRF)Week 08 · Classes 22–24+
- Class 22 · Dense vs Sparse Retrieval Dense vector embeddings vs sparse keyword algorithms (BM25, SPLADE) and vocabulary mismatch.
- Class 23 · Reciprocal Rank Fusion (RRF) RRF mathematical formulations, dynamic k tuning, and convex score combination techniques.
- Class 24 · Cross-Encoder Re-rankers Cohere Rerank, BGE-Reranker, bi-encoder vs cross-encoder cost-latency trade-offs, and top-k filtering.
09Vector Database Internals at ScaleWeek 09 · Classes 25–27+
- Class 25 · Vector Indexing Math Flat vs IVF (Inverted File) vs HNSW (Hierarchical Navigable Small World); tuning M and efConstruction.
- Class 26 · PgVector in Enterprise Production HNSW indexing in PostgreSQL, IVFFlat, halfvec (FP16), vacuuming, and memory allocation.
- Class 27 · Milvus & Qdrant Enterprise Clustering Partitioning, payload indexing, write-ahead logging (WAL), and distributed query coordination.
10Multi-Tenancy & Row-Level Security (RLS)Week 10 · Classes 28–30+
- Class 28 · Multi-Tenancy Isolation Models Schema-per-tenant vs collection-per-tenant vs metadata-filtered partitions.
- Class 29 · Row-Level Security (RLS) in RAG Syncing Active Directory/OAuth roles into vector payloads and deterministic filter enforcement.
- Class 30 · Context-Window Optimizations Selective token pruning, contextual compression, summarization trees, and 'lost-in-the-middle' mitigation.
11GraphRAG & Relational Knowledge RetrievalWeek 11 · Classes 31–33+
- Class 31 · Limits of Semantic Similarity Why pure vectors fail at multi-hop reasoning; introduction to Knowledge Graphs for RAG.
- Class 32 · Graph Extraction Pipelines Automated entity-relation extraction using LLMs into Neo4j/Amazon Neptune.
- Class 33 · Hybrid GraphRAG Implementation Combining vector search with graph traversal queries (Cypher + Vector similarity).
12Production Milestone Project 2Week 12 · Classes 34–36+
- Class 34 · Production Lab 1 Architecting a multi-tenant enterprise RAG pipeline supporting 100K+ complex PDF/DOCX files.
- Class 35 · Production Lab 2 Implementing GraphRAG + RRF reranking with strict metadata RBAC validation.
- Class 36 · Evaluation & Defense Benchmarking retrieval against golden datasets using Hit Rate@K and MRR; architectural defense.
Module 03 · Weeks 13–18
Stateful Agentic Systems & Model Context Protocol (MCP)
Focus: Complex multi-agent coordination, deterministic execution graphs, and tool protocols.
13Deep Dive into State Graphs with LangGraphWeek 13 · Classes 37–39+
- Class 37 · Cyclic Execution Graphs Beyond linear DAGs: cyclical execution loops, conditional routers, and branch merges in LangGraph.
- Class 38 · State Reducer Architecture State management, channel reducers, typed schemas, and immutability patterns in long-running agents.
- Class 39 · Production Checkpointing PostgreSQL and Redis state checkpointers for fault recovery, session continuity, and pause/resume.
14Human-in-the-Loop & Time-Travel DebuggingWeek 14 · Classes 40–42+
- Class 40 · Human-in-the-Loop (HITL) Workflows Manual approval gates, review queues, webhook interruptions, and escalation thresholds.
- Class 41 · State Editing & Time-Travel Rewinding graph execution, editing past agent state, branching workflows, and replaying history.
- Class 42 · Subgraphs & Agent Decomposition Modular agent architecture, nested subgraphs, and context isolation across agent teams.
15Multi-Agent Collaboration & SupervisionWeek 15 · Classes 43–45+
- Class 43 · Multi-Agent Topologies Supervisor patterns, hierarchical command trees, and choreographed peer-to-peer collaboration.
- Class 44 · Shared vs Isolated Agent Memory Working memory scratchpads, cross-agent message passing, and token window conservation.
- Class 45 · Deadlock & Loop Circuit Breakers Circular dependency detection, maximum iteration enforcement, and token exhaustion circuit breakers.
16Model Context Protocol (MCP) in ProductionWeek 16 · Classes 46–48+
- Class 46 · MCP Architecture & Specifications Client-server protocol specification, JSON-RPC messaging, stdio vs Server-Sent Events (SSE) transports.
- Class 47 · Building Production MCP Servers Exposing databases, microservices, and file management tools safely through MCP endpoints.
- Class 48 · Building Enterprise MCP Hosts Dynamic tool discovery, access permission enforcement, and authentication propagation.
17Determinism & Schema GovernanceWeek 17 · Classes 49–51+
- Class 49 · Enforcing Deterministic Outputs Pydantic v2 schemas, Instructor library, automatic self-correction loops, and structured validation.
- Class 50 · Agent Evaluation Frameworks Trajectory evaluation, tool-selection precision/recall, and step-efficiency scoring.
- Class 51 · Agent Sandboxing & Execution Safety Running untrusted agent-generated code inside isolated microVMs (gVisor/Firecracker/Docker).
18Production Milestone Project 3Week 18 · Classes 52–54+
- Class 52 · Production Lab 1 Building a stateful multi-agent automated code review and QA system using LangGraph and custom MCP servers.
- Class 53 · Production Lab 2 Implementing PostgreSQL checkpointers, HITL approval webhooks, and sandboxed code execution.
- Class 54 · Review & Chaos Testing Chaos testing agent recovery under node failure and comprehensive code review.
Module 04 · Weeks 19–24
Domain Adaptation, Supervised Fine-Tuning & Alignment
Focus: When prompting and RAG fall short: custom model training, PEFT, and preference alignment.
19Dataset Engineering & Synthetic GenerationWeek 19 · Classes 55–57+
- Class 55 · Fine-Tuning Decision Matrix When to Fine-Tune vs RAG vs Prompt Engineering: technical boundaries, latency, and ROI modeling.
- Class 56 · Enterprise Dataset Curation Data cleansing, MinHash LSH deduplication, decontamination, and format standardization.
- Class 57 · Synthetic Data Generation LLM-as-a-Judge pipelines for dataset expansion, rejection sampling, and diversity filtering.
20Parameter-Efficient Fine-Tuning (PEFT)Week 20 · Classes 58–60+
- Class 58 · LoRA Mathematical Foundations Low-Rank Adaptation math, intrinsic dimensionality, rank r, and scaling factor α.
- Class 59 · QLoRA & 4-Bit Quantization NormalFloat4 (NF4), double quantization, and paged optimizers for memory-constrained training.
- Class 60 · Accelerated Fine-Tuning with Unsloth Custom CUDA kernels, gradient checkpointing, and high-throughput SFT using Hugging Face TRL.
21Training Infrastructure & Supervised Fine-Tuning (SFT)Week 21 · Classes 61–63+
- Class 61 · Hyperparameter Tuning for LLMs Cosine learning rate schedules with warmup, weight decay, batch sizes, and gradient accumulation.
- Class 62 · Distributed SFT Infrastructure PyTorch FSDP (Fully Sharded Data Parallel) and DeepSpeed ZeRO-1/2/3 strategies for multi-GPU training.
- Class 63 · Training Monitoring & Diagnostics Loss curve analysis, perplexity tracking, gradient norm monitoring, and catastrophic forgetting prevention.
22Alignment: DPO & ORPOWeek 22 · Classes 64–66+
- Class 64 · Preference Data Architecture Constructing paired preference datasets (chosen vs rejected responses) for domain alignment.
- Class 65 · Direct Preference Optimization (DPO) Mathematical derivation of DPO vs RLHF/PPO; eliminating the reward model overhead.
- Class 66 · ORPO & Safe Alignment Odds Ratio Preference Optimization and instruction-following alignment without degradation.
23Merging, Exporting & Multi-LoRA ServingWeek 23 · Classes 67–69+
- Class 67 · Weight Merging & Precision Handling Merging LoRA adapters into base weights, 16-bit de-quantization, and GGUF/AWQ export.
- Class 68 · Dynamic Multi-LoRA Serving Configuring vLLM to serve dozens of specialized LoRA adapters concurrently on a single base model.
- Class 69 · Comprehensive Model Benchmarking Running task-specific evaluation suites, MMLU regression checks, and domain accuracy metrics.
24Production Milestone Project 4Week 24 · Classes 70–72+
- Class 70 · Production Lab 1 Fine-tuning an open-weights model on proprietary structured domain documents using QLoRA.
- Class 71 · Production Lab 2 Aligning the model with DPO, validating benchmark regressions, and deploying as a multi-LoRA endpoint on vLLM.
- Class 72 · Model Evaluation & Sign-Off Final evaluation against unseen validation datasets and checkpoint auditing.
Module 05 · Weeks 25–28
LLMOps, Continuous Evaluation, Security & Observability
Focus: Industrializing AI deployments with enterprise security, telemetry, and automated CI/CD.
25Continuous Evaluation-Driven Development (EDD)Week 25 · Classes 73–75+
- Class 73 · Synthetic Benchmark Suites Creating representative golden ground-truth test datasets for continuous evaluation.
- Class 74 · RAG Evaluation Frameworks Deep dive into Ragas and DeepEval: Faithfulness, Answer Relevance, Context Precision, and Recall.
- Class 75 · Custom LLM Judges & CI Evals Designing calibrated LLM-as-a-Judge evaluators and integrating them into automated test suites.
26Distributed Observability & TelemetryWeek 26 · Classes 76–78+
- Class 76 · OpenTelemetry for GenAI Semantic conventions for GenAI, spans, traces, and metadata propagation across microservices.
- Class 77 · Full-Stack Tracing Tools Deploying LangSmith and Arize Phoenix: tracing nested tool executions, latency sinks, and error rates.
- Class 78 · Production Metric Dashboards Tracking token usage, cost-per-user, TTFT, TPOT, and latency percentiles in Prometheus and Grafana.
27AI Security & OWASP Top 10 for LLMsWeek 27 · Classes 79–81+
- Class 79 · Prompt Injection Defense Detecting and preventing direct/indirect prompt injection, jailbreaks, and payload smuggling.
- Class 80 · Guardrail Frameworks Implementing NeMo Guardrails, Llama Guard, and regex/PII masking engines in the inference path.
- Class 81 · Model Supply Chain Security Safetensors validation, poisoned dataset prevention, and API key token rotation architectures.
28LLMOps CI/CD Pipelines & Cost GovernanceWeek 28 · Classes 82–84+
- Class 82 · Automated Prompt CI/CD GitHub Actions workflows for prompt regression testing and threshold validation before deployment.
- Class 83 · Semantic Caching & Cost Routing Redis semantic caching, prompt caching, model cascades (routing queries by complexity), and budget throttles.
- Class 84 · Canary & Blue/Green Deployments Zero-downtime routing strategies for new prompt templates and fine-tuned model versions.
Module 06 · Weeks 29–32
Capstone Architecture & Defense
Focus: End-to-end delivery of an enterprise-grade AI system with strict SLA and cost boundaries.
29System Scoping, SLA Definition & Architecture BlueprintWeek 29 · Classes 85–87+
- Class 85 · Non-Functional Scoping Defining strict latency floors (P95/P99), dollar cost ceilings, and uptime targets.
- Class 86 · Retrieval & Ingestion Blueprint Architecting high-throughput data pipelines and hybrid vector/graph schemas.
- Class 87 · Serving & High Availability Blueprint Architecting GPU clusters, fallback routing, and checkpointer resilience.
30Implementation & Infrastructure ProvisioningWeek 30 · Classes 88–90+
- Class 88 · Core Engine Deployment Deploying the multi-agent execution core and low-latency inference cluster.
- Class 89 · Retrieval & Tool Integration Integrating RLS-compliant hybrid retrieval and custom MCP microservices.
- Class 90 · Telemetry & Guardrail Hooks Wiring OpenTelemetry distributed tracing and bidirectional guardrails.
31Stress Testing, Chaos Engineering & HardeningWeek 31 · Classes 91–93+
- Class 91 · Chaos Testing & Node Failures High-concurrency load testing and GPU node failure injection under production loads.
- Class 92 · Adversarial Red-Teaming Penetration testing with automated jailbreak payloads and indirect injection vectors.
- Class 93 · Cost & Performance Tuning Maximizing semantic cache hit rates and tuning continuous batching parameters.
32Production Rollout & Architectural DefenseWeek 32 · Classes 94–96+
- Class 94 · Canary Deployment Execution Canary traffic routing, real-time drift monitoring, and automated rollback validation.
- Class 95 · Boardroom Technical Defense Presenting and defending system architecture, trade-offs, and metrics to senior engineering evaluators.
- Class 96 · Final Audit & Graduation Comprehensive code audit, architecture sign-off, and program completion.
HOW YOU LEARN
Build it. Ship it.
Defend it.
Three hands-on classes a week, with production labs and architecture reviews that mirror how senior engineering teams work.
See the curriculum01◷
Right-sized sessions
Every session runs 1.5–2 hours, 3 classes a week.
02▦
32 weeks, 96 classes
About 7.5–8 months of interactive, hands-on production sessions.
03⌁
Production milestone projects
Modules 01–04 end with production labs, load testing or evaluation, and a milestone sign-off.
04◎
Architectural defense
Code reviews, failure-recovery drills, chaos testing and a final boardroom technical defense.
WHO IT’S FOR
Built for engineers
who already ship.
For software and data engineers with 2–8 years of experience who need to deploy robust, cost-effective and observable AI systems at scale.
Data engineers
Turn data pipelines into multi-tenant retrieval, GraphRAG and fine-tuning datasets.
Senior & lead engineers
Own the architecture: SLAs, cost ceilings, observability, security and production rollout.
For experienced engineers: 2–8 years in software or data engineering. Newer to AI? Start with the AI Engineer program.
BATCHES
Enrol in
AI Engineer Advanced.
Talk to us about upcoming batch dates, fees and payment options.
Reserve your seat
Share your details and our team will get back to you with batch dates, fees and payment options.
- Duration32 weeks (7.5–8 months)
- Classes96 · 3 per week
- Session length1.5–2 hours
- Experience2–8 years
READY WHEN YOU ARE
Let’s build something
production-grade.
Tell us what you want to learn and we’ll help you explore the program.