IICTI Certified Large Language Model Engineer

Holders demonstrate that they can apply and evaluate transformer internals and efficient model architectures, pretraining data pipelines, tokenizers and scaling laws, distributed training and parameter-efficient fine-tuning and post-training alignment and reasoning optimisation, plus 8 further areas, in realistic professional scenarios, as defined in the AQSCERT scheme for Large Language Model Engineer.

Who it is for: Professionals working, or preparing to work, in artificial intelligence roles.

1. 01Transformer internals and efficient model architectures6% of the exam

   • Evaluate architectural choices for quality, context length, throughput and memory trade-offs.
   • Scope: Attention variants (MHA, GQA, MLA), FlashAttention, RoPE and context extension (YaRN), mixture-of-experts routing and load balancing, state-space and hybrid architectures, KV cache behaviour.
2. 02Pretraining data pipelines, tokenizers and scaling laws6% of the exam

   • Design data and tokenization pipelines and compute budgets grounded in scaling-law evidence.
   • Scope: BPE and Unigram tokenizer design, multilingual coverage, MinHash deduplication, quality and safety filtering, benchmark contamination, data mixture optimisation, compute-optimal scaling laws.
3. 03Distributed training and parameter-efficient fine-tuning6% of the exam

   • Architect distributed training and fine-tuning runs and diagnose convergence and stability failures.
   • Scope: FSDP and ZeRO, tensor, pipeline and expert parallelism, BF16/FP8 mixed precision, gradient checkpointing, training instability diagnosis, LoRA, QLoRA and DoRA, continued pretraining.
4. 04Post-training alignment and reasoning optimisation6% of the exam

   • Design post-training pipelines that improve helpfulness, reasoning and safety while controlling reward hacking and regressions.
   • Scope: Supervised fine-tuning, reward modelling, RLHF with PPO, DPO and variants, GRPO and RL with verifiable rewards, RLAIF and constitutional methods, reward hacking and overoptimisation.
5. 05High-throughput inference serving and optimisation10% of the exam

   • Architect inference serving that meets latency, throughput and cost targets and diagnose serving bottlenecks.
   • Scope: vLLM PagedAttention, SGLang and TensorRT-LLM, continuous batching, prefix caching, speculative decoding, quantization (GPTQ, AWQ, FP8), disaggregated prefill and decode, GPU capacity planning.
6. 06Retrieval-augmented generation architecture16% of the exam

   • Design and tune RAG systems and diagnose retrieval versus generation failures using targeted metrics.
   • Scope: Chunking and contextual retrieval, hybrid BM25 and dense search, rerankers, vector database indexing (HNSW, IVF-PQ), GraphRAG, citation grounding, RAG evaluation (faithfulness, context recall).
7. 07Agentic systems, tool use and context engineering8% of the exam

   • Architect reliable agentic workflows with bounded autonomy, robust tool integration and recoverable failure modes.
   • Scope: Function calling and structured outputs, Model Context Protocol, planning and reflection loops, multi-agent orchestration, memory and context window management, sandboxed execution, human-in-the-loop controls.
8. 08LLM evaluation methodology and benchmarking10% of the exam

   • Design evaluation strategies that produce statistically sound, decision-ready evidence of model and system quality.
   • Scope: Task-specific eval suites, LLM-as-judge calibration and bias, pairwise preference and Elo, contamination detection, statistical significance and variance, regression gating in CI, online A/B testing.
9. 09LLM security and adversarial robustness10% of the exam

   • Evaluate LLM applications for adversarial risk and design layered defences proportionate to agent privileges and data exposure.
   • Scope: OWASP Top 10 for LLM Applications 2025, direct and indirect prompt injection, data exfiltration via tools, jailbreak resistance, MITRE ATLAS, guardrail and classifier layers, AI red teaming governance.
10. 10LLMOps, observability and cost engineering11% of the exam

    • Lead production LLM operations with observability, controlled releases and cost-per-task optimisation.
    • Scope: OpenTelemetry GenAI semantic conventions, tracing of prompts and tool calls, prompt and model versioning, drift and quality monitoring, semantic caching, token cost attribution, model routing and fallbacks.
11. 11Multimodal and long-context model engineering5% of the exam

    • Evaluate and integrate multimodal and long-context capabilities against task requirements and their failure modes.
    • Scope: Vision-language model architectures, image and audio encoders, document understanding, long-context retrieval and degradation (needle and multi-hop tests), context compression.
12. 12Responsible AI governance and regulatory compliance6% of the exam

    • Lead governance of LLM development and deployment that satisfies regulatory, documentation and risk management obligations.
    • Scope: EU AI Act GPAI obligations and Code of Practice, ISO/IEC 42001, NIST AI RMF and Generative AI Profile (AI 600-1), model and system cards, training data provenance and copyright, privacy in training data.

Question types

Single choice 100%

Delivery

Online (remote proctored), Test centre, On-site (corporate)

Retakes

Up to three attempts within any 12-month period. A 14-day waiting period applies after a first unsuccessful attempt and 30 days after a second. Each retake requires a new booking and fee. Retakes draw a different randomised form from the question bank.

Provisional cut score — psychometric validation is required before live delivery. Blueprint v1.0.

1. 1

   Register

   Create a free account and start the programme.

2. 2

   Practice

   Online practice test (coming soon).

3. 3

   Training

   Optional

   Online course or classroom — or skip it.

4. 4

   Certification

   Independent exam and certification decision.

Training is optional and does not by itself confer certification. Find a training partner