EPGC IN DATA SCIENCE & APPLIED AI - Collaboration with AWS

Learning objective

Build calibrated classification models and optimise decision thresholds against asymmetric business cost matrices to maximise return on investment. Evaluate models on imbalanced enterprise data, calibrate predicted probabilities, and deploy real-time scoring on Amazon SageMaker Serverless Inference.

Topics:

• Conditional probability and Bayes' theorem
• Precision-recall tradeoffs on skewed datasets and business utility cost matrices
• Logistic Regression, the sigmoid function and log-loss
• The confusion matrix: precision, recall, F1-score and Precision-Recall Area Under Curve (PR-AUC)
• Why accuracy is misleading on imbalanced data, and class-weighted loss
• Receiver Operating Characteristic (ROC) and precision-recall curves across threshold intervals
• Asymmetric cost matrices and the probability threshold that maximises net return
• Probability calibration: Platt Scaling, Isotonic Regression and Brier scores
• Reliability calibration curves for detecting overconfident models
• Amazon SageMaker Serverless Inference endpoints for on-demand scoring
• Batch risk alerting for retention workflows

Hands-on

• scikit-learn
• NumPy
• Matplotlib
• Seaborn
• Jupyter Notebook
• Amazon SageMaker (Serverless Inference)
• Amazon S3
• AWS SDK for Python (Boto3)

Project

B2B Churn Early-Warning Engine

How to detect high-risk customer churn and optimise retention return on investment. Built Cost-Utility Optimized Churn Decision Service with an alert API.

Learning objective

Train distributed gradient boosted tree ensembles, optimise hyperparameters with Bayesian search and govern model lineage in production registries. Run managed XGBoost training on AWS, tune efficiently with Automatic Model Tuning, and audit models for bias before registering approved versions.

Topics:

• Information theory basics: entropy versus Gini impurity
• Bagging versus boosting, and hyperparameter search spaces
• Recursive binary splitting and decision tree overfitting
• Variance reduction through bagging with Random Forests
• Gradient boosting: fitting trees to gradient residuals with XGBoost and LightGBM
• Distributed training with the Amazon SageMaker managed XGBoost container and early stopping
• Training and validation objective loss curves
• Grid search, random search and Bayesian Optimisation with Gaussian Process surrogate models
• SageMaker Automatic Model Tuning (AMT) hyperparameter ranges: max_depth, eta and subsample
• Bias auditing with SageMaker Clarify: disparate impact and feature attributions
• Model Cards, metadata lineage and versioned model packages in SageMaker Model Registry

Hands-on

• scikit-learn
• XGBoost
• LightGBM
• Matplotlib
• Amazon SageMaker (Managed XGBoost)
• Amazon SageMaker Automatic Model Tuning (AMT)
• Amazon SageMaker Clarify
• Amazon SageMaker Model Registry
• Amazon S3
• AWS SDK for Python (Boto3)

Project

Fleet Logistics Delay Predictor

How to forecast delivery delays with XGBoost and audit model fairness. Deliver production built on Governed Fleet Delay Model Package & Audit Model Card.

Learning objective 

Perform dimensionality reduction, discover operational behaviour clusters, and score anomalies on continuous telemetry in real time. Apply Principal Component Analysis to high-dimensional sensor data, validate cluster quality statistically, and detect equipment anomalies before failure. 

Topics: 

• Geometric distance formulations: Euclidean, Manhattan and Cosine 

• The mathematical necessity of feature scaling, and orthogonal variance projection 

• Principal Component Analysis (PCA): covariance matrices, eigenvectors and eigenvalues 

• Explained variance ratio and scree plots 

• K-Means clustering, centroid update mathematics and the Voronoi partition 

• Cluster validation with the Elbow Method (inertia) and Silhouette Analysis 

• Operational behaviour profiling across machine states 

• Unsupervised anomaly detection with Isolation Forests and random space partitioning 

• Anomaly scoring from average path lengths 

• Real-time anomaly scoring on continuous streams with SageMaker Random Cut Forest 

Hands-on 

• scikit-learn 

• NumPy 

• Matplotlib 

• Seaborn 

• Jupyter Notebook 

• Amazon SageMaker (Random Cut Forest) 

• Amazon SageMaker 

• Amazon S3 

• AWS SDK for Python (Boto3) 

Project 

IoT Sensor Clustering & Anomaly Engine 

How to detect machine failure in real time across 40+ sensor streams. Building Continuous Machine Health Monitoring & Anomaly Service.

Learning objective 

Decompose temporal demand patterns, train recurrent deep learning sequence models and generate risk-aware probabilistic forecasts across many series. Test stationarity statistically, build LSTM architectures in PyTorch, and produce quantile forecasts with Amazon SageMaker DeepAR. 

Topics: 

• Classical time-series components: trend, seasonality, stationarity and autocorrelation 

• Feedforward artificial neural network mechanics 

• Time-series decomposition into trend, seasonal and residual components 

• Augmented Dickey-Fuller (ADF) stationarity testing and lag differencing 

• Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) analysis 

• Recurrent Neural Networks (RNNs), the vanishing gradient problem and Long Short-Term Memory (LSTM) gating 

• Sliding-window sequence feature matrices and multi-step forecast evaluation 

• Probabilistic forecasting: predicting distributions rather than point estimates 

• SageMaker DeepAR autoregressive architecture and cross-learning across thousands of series 

• Quantile forecasts (P10, P50, P90) and DeepAR JSONLines formatting with dynamic features 

• Weighted quantile loss (wQL) evaluation and safety stock recommendations 

AWS Self-Paced Content 

• Getting Started with Amazon Comprehend 

• Amazon Rekognition Getting Started 

• Cloud Financial Management: FinOps for GenAI 

Hands-on 

• statsmodels 

• pandas 

• Matplotlib 

• Jupyter Notebook 

• PyTorch 

• scikit-learn 

• Amazon SageMaker (DeepAR) 

• Amazon S3 

• AWS SDK for Python (Boto3) 

Project 

Multi-SKU Probabilistic Demand Forecaster 

Predict retail inventory peaks using DeepAR forecasting, and export a Risk-Aware Probabilistic SKU Replenishment Engine. 

Learning objective 

Build multimodal document extraction pipelines, generate dense vector embeddings and power semantic search engines. Drive foundation models programmatically for deterministic structured output, parse complex documents with Amazon Textract, and run vector search on Amazon OpenSearch Service.

Topics: 

• Text tokenisation and vector space models, from TF-IDF to dense embeddings 

• Self-attention transformer intuition 

• Transformer architecture: multi-head self-attention and context windows 

• Sampling parameters: Temperature and Top-P for controlling generation entropy 

• Enforcing deterministic JSON output with the Amazon Bedrock Converse API and Pydantic schemas 

• Multimodal document AI: layout parsing, key-value pair detection and tabular extraction 

• Parsing scanned, skewed, multi-column invoices with Amazon Textract 

• Validating extracted financial totals against business rules using foundation models 

• Lexical keyword matching versus dense vector embeddings 

• 1536-dimensional Amazon Titan embeddings 

• OpenSearch vector indexes with the HNSW engine and k-Nearest Neighbour (k-NN) / Approximate Nearest Neighbor (ANN) search 

• Hybrid semantic search across an enterprise product catalogue 

AWS Self-Paced Content 

• Amazon Textract Getting Started 

• Amazon Transcribe Getting Started 

Hands-on 

• Amazon Bedrock (Converse API) 

• Amazon Bedrock (Titan Embeddings) 

• Amazon Textract 

• Amazon OpenSearch Service 

• Pydantic 

• Python 

• Jupyter Notebook 

• AWS SDK for Python (Boto3) 

Project 

Multimodal Document AI & Semantic Search 

Parse complex supplier invoices and enable vector search and deliver a Multimodal Invoice Parser and Semantic Product Search API.

Learning objective 

Build enterprise Retrieval-Augmented Generation knowledge engines, orchestrate autonomous multi-step reasoning agents with tool calling, and enforce safety guardrails. Ground answers in company documents with citations, execute real backend actions through Bedrock Agents, and defend against prompt injection.

Topics: 

• Retrieval-Augmented Generation (RAG) components and context chunking strategies 

• Prompt engineering patterns and the large language model vulnerability taxonomy 

• Vector retrieval and context injection into prompts for grounded, cited answers 

• Amazon Bedrock Knowledge Bases with automated chunking and OpenSearch Serverless indexing 

• The ReAct framework: reasoning combined with acting 

• Function calling and OpenAPI 3.0 tool schemas 

• Amazon Bedrock Agents with AWS Lambda action groups that execute real business actions 

• Prompt injection and data extraction attacks against unguarded models 

• Jailbreak detection, contextual grounding checks and denied-topic filtering 

• Masking Personally Identifiable Information (PII) such as social security and credit card numbers with Amazon Bedrock Guardrails 

• End-to-end multi-turn agent testing and validation traces 

Hands-on 

• Amazon Bedrock Knowledge Bases 

• Amazon Bedrock Agents 

• Amazon Bedrock Guardrails 

• Amazon OpenSearch Serverless 

• AWS Lambda 

• OpenAPI 3.0 Schemas 

• Amazon S3 

• Jupyter Notebook 

• AWS SDK for Python (Boto3) 

Project 

Autonomous IT Support Concierge 

Deploy Bedrock RAG agents with safety guardrails to resolve tickets. And deliver a Secure Autonomous IT Concierge Agent.

Learning objective 

Engineer low-latency real-time serving endpoints, configure elastic autoscaling, detect data and concept drift, and execute zero-downtime rollouts. Package models behind custom inference handlers, scale them under traffic surges, and validate new versions on live traffic before release.

Topics: 

• RESTful API concepts and containerisation principles: Docker images versus runtimes 

• Model performance degradation and distribution drift concepts 

• Custom inference handlers: model_fn, input_fn, predict_fn and output_fn 

• Multi-worker concurrency and SageMaker Real-Time Endpoints 

• Benchmarking inference latency under concurrent load 

• Target Tracking autoscaling on SageMakerVariantInvocationsPerInstance 

• Multi-Model Endpoints (MME) hosting several models on a shared container fleet 

• Data drift and concept drift: Kolmogorov-Smirnov tests and the Population Stability Index (PSI) 

• SageMaker Model Monitor data capture, statistical baselines and CloudWatch violation alarms 

• SageMaker Production Variants for live A/B traffic splitting 

• Shadow deployment and zero-downtime traffic shifting to a candidate model 

AWS Self-Paced Content 

• AWS Cost Explorer Getting Started 

• AWS Budgets Getting Started 

• Introduction to Amazon Virtual Private Cloud (VPC) 

Hands-on 

• Amazon SageMaker (Real-Time Endpoints) 

• Amazon SageMaker (Autoscaling & Multi-Model Endpoints) 

• Amazon SageMaker Model Monitor 

• Amazon SageMaker Production Variants 

• Amazon CloudWatch 

• Docker 

• Amazon S3 

• AWS SDK for Python (Boto3) 

Project 

Zero-Downtime Serving Mesh & MLOps 

How to deploy real-time SageMaker endpoints with live drift alarms. Delivering a Resilient Zero-Downtime Serving Mesh with drift detection.

Learning objective 

Design, implement, deploy and defend an end-to-end production AI platform before executive leadership and an industry panel. Blueprint the system against the AWS Well-Architected Framework ML Lens, build the automated data and training pipeline, deploy the serving and GenAI layer with guardrails and drift monitoring, then defend it live.

Topics:

• Milestone 1: constructing the automated data platform and model training engine. We will ingest raw datasets into S3, execute PySpark ETL and DQDL quality validation in Glue, train models via SageMaker Script Mode, and register approved versions in SageMaker Model Registry.

• Milestone 2: building the production serving and intelligence layer. We will deploy autoscaling SageMaker real-time endpoints, integrate Amazon Bedrock GenAI agents with tool calling, enforce Bedrock Guardrails safety, and configure Model Monitor drift alarms. Live end-to-end system walkthrough and defence of cost, safety and resilience trade-offs

Hands-on

• AWS Well-Architected Tool

• Diagrams /Amazon Draw.io

• AWS Architecture Center

• Amazon S3

• AWS Glue

• Amazon SageMaker

• Amazon SageMaker Model Registry

• Amazon Bedrock

• Amazon Bedrock Guardrails

• Amazon CloudWatch

• GitHub

• AWS SDK for Python (Boto3)

Project

Apex Enterprise AI Capstone Defence

Architect, harden and defend an end-to-end cloud AI platform. Deliver a live demonstration from raw data ingestion to real-time inference and autonomous agent interactions and present a Production Enterprise AI Platform for executive sign-off. Evaluated on system architectural resilience (40%), GenAI safety and real-time action groups (30%), and MLOps automation and continuous drift monitoring (30%).