Course Objective & Scope
A Transformer that scores high F1 in a notebook quickly runs into variable sequence lengths, GPU out-of-memory spikes, PII leakage, shifting user vocabulary and runaway cloud bills in production. This engineering programme covers optimising encoder and seq2seq models with Hugging Face Optimum, ONNX Runtime and TensorRT (FP16/INT8), eliminating padding waste via length bucketing and dynamic batching, and serving with FastAPI and Triton Inference Server.
For production reliability, participants implement PII redaction with Microsoft Presidio, behavioural regression testing (CheckList for negation, typos and entity swaps), embedding-drift monitoring and cost-aware cascade routing from rules/compact encoders to heavier models.
Who Should Attend?
- ML and MLOps/LLMOps engineers deploying NLP and language models to production
- Backend and platform engineers building high-throughput text processing and search services
- Technical leads responsible for latency, unit cost, privacy and reliability of NLP pipelines
- Prerequisite: Python, basic Docker/API knowledge and experience with Transformer models
Key Takeaways
- Accelerate Transformer inference 3–6x and shrink memory footprints with Hugging Face Optimum, ONNX Runtime and TensorRT
- Cut p99 latency on variable-length text using length bucketing and token-budget dynamic batching
- Build GDPR/KVKK-compliant PII detection and redaction pipelines with Microsoft Presidio and custom Turkish recognisers
- Gate model releases in CI/CD with CheckList behavioural, negation and typo-invariance test suites
- Design confidence-routed model cascades and monitor live traffic for vocabulary and embedding drift
Curriculum
Production NLP Architecture and Cascade Routing
- Synchronous low-latency APIs vs asynchronous stream/batch text pipelines
- Confidence-routed cascades: Regex/rules -> compact encoder -> heavy model/LLM
Model Optimisation: Optimum, ONNX and Quantisation
- Graph export and operator fusion with Hugging Face Optimum, ONNX and TensorRT
- Dynamic and static INT8/FP16 quantisation for CPU (AVX/VNNI) and GPU targets
High-Throughput Serving and Dynamic Batching
- Solving the variable-length padding bottleneck via length bucketing and token batching
- Serving with Triton Inference Server / FastAPI, semantic caching and load testing
PII Redaction, Safety and Behavioural CI/CD Testing
- PII anonymisation with Microsoft Presidio and custom local entity recognisers
- CheckList suites: negation sensitivity, typo robustness and entity-swap invariance
Observability, Text Drift and Active Feedback Loops
- Monitoring subword fertility, sequence length shifts and embedding drift in production
- Routing low-confidence slices to human review and automated retraining pipelines
Other Courses in This Category
-
Turning unstructured documents into structured tables and JSON: hierarchical and multi-label classification, GLiNER and Transformer NER, relation and event extraction, layout-aware PDF parsing and schema-validated extraction.
- Live Course
- 5 modules
-
Enterprise search beyond keyword matching: Sentence-Transformers (bi-encoders), multilingual embeddings, BM25/SPLADE + dense hybrid retrieval, Reciprocal Rank Fusion, Cross-Encoder reranking, Qdrant/pgvector and NDCG@k evaluation.
- Live Course
- 5 modules
-
Deep Learning with PyTorch
Comprehensive deep learning training with PyTorch 2.x covering tensor operations, autograd, CNNs, Transformers, LLM fine-tuning (LoRA/QLoRA), diffusion models, GNNs, and production deployment.
- Live Course
- 12 modules
-
Neural networks, forward pass, backpropagation, activation and loss functions, AdamW, dropout and batch normalisation, giving engineers the skills to build and train their first deep learning models in PyTorch.
- Live Course
- 5 modules
-
Stable, large-scale deep learning covering BF16/FP16 mixed precision, multi-GPU training with DDP/FSDP, self-supervised and contrastive pretraining, metric learning, knowledge distillation and GPU profiling.
- Live Course
- 5 modules
-
From ResNet and ConvNeXt to Vision Transformers (ViT), Swin, hybrid backbones, State Space Models (Mamba), Mixture-of-Experts (MoE) and multimodal fusion: selecting and designing architectures for your problem and hardware.
- Live Course
- 5 modules
Get in touch
