Telstra Data & AI

Delivered multiple applied machine learning and AI initiatives spanning fraud detection, LLM optimisation, MLOps infrastructure, and responsible AI, supporting enterprise-scale data platforms and ethical AI governance across Telstra’s Data & AI organisation.

Fraud Detection – Call Transcript Classification

  • Developed an end-to-end machine learning pipeline to automatically detect fraudulent behaviour in call transcripts, improving fraud identification accuracy and reducing manual review time by over 60%.

  • Framed and scoped the business problem by analysing how fraudulent calls impact operational efficiency, compliance, and customer trust.

  • Engineered and prepared large volumes of text data (hundreds of thousands of transcripts), handling missing values, noise, and class imbalance to ensure robust model performance.

  • Applied multiple classification algorithms (Logistic Regression, RandomForest, XGBoost) and optimised performance using hyperparameter tuning to minimise false positives.

  • Addressed key challenges such as overfitting, data wrangling, and memory constraints through efficient feature selection, stratified sampling, and cross-validation.

  • Enhanced model interpretability using RAKE and ELI5 to generate explainable results for non-technical stakeholders.

  • Delivered insights and model results via interactive dashboards and visualisations (Matplotlib, Seaborn, ipywidgets), enabling the business to monitor fraud patterns and refine prevention strategies.

  • Tools and libraries: Python, Pandas, NumPy, Scikit-learn, SpaCy, RAKE, ELI5, ipywidgets.

QAI – Intelligent Call Transcript Triage

  • Developed a hybrid classification pipeline combining XGBoost and Logistic Regression to triage call transcripts before GPT analysis, cutting LLM processing time by ~80% and significantly reducing compute costs.

  • Business goal: optimise compliance detection while reducing the dataset processed by LLMs to under 25% of its original size.

  • Processed and modelled over 1.5 million call transcripts across three compliance categories (DL, PA, AU), targeting high recall to minimise missed compliance issues.

  • Preprocessed and vectorised text data using TF-IDF with lemmatisation, stop-word removal, and n-gram extraction to transform text into structured numerical vectors.

  • Addressed extreme class imbalance (positives < 5%) through undersampling, threshold tuning, and recall-driven optimisation, achieving recalls of 0.94–0.98 across models.

  • Implemented an ensemble voting strategy that combined model strengths — Logistic Regression for positives, XGBoost for negatives — with adjustable thresholds for optimal balance.

  • Applied extensive hyperparameter optimisation using RandomisedSearchCV and GridSearchCV.

  • Enhanced GPT classification accuracy by refining query design and reducing irrelevant inputs.

  • Impact: reduced daily transcript processing from 8 hours to under 2, enabling faster compliance insights and substantial cost savings.

  • Tools and libraries: Python, Scikit-learn, XGBoost, TF-IDF, Pandas, NumPy.

MLOps & Orchestrator – Data Platform Engineering

  • Built a scalable data platform with full MLOps and LLMOps capabilities, including CI/CD pipelines, continuous monitoring, and automated orchestration for ML model deployment and maintenance.

  • Developed orchestrator workflows using Airflow, Docker, and Postgres on Ubuntu/WSL, enabling reproducible and auditable data processing pipelines.

  • Integrated feature quality monitoring dashboards in Amazon Quicksight to track data ingestion and detect drift across model features.

  • Enhanced automated Python validation scripts to dynamically identify and flag data anomalies, improving input reliability prior to model training.

  • Built and refined New Relic dashboards to monitor data quality across Quantium–Telstra data models, reducing issue detection time and improving data confidence.

  • Aligned platform design with Azure DevOps CI/CD standards and implemented Function App–based automation, standardising DAG execution and improving runtime consistency.

  • Impact: streamlined model delivery, improved data reliability, and established a foundation for continuous monitoring and scalable AI model operations across the organisation.

Risk & Ethics – AI Risk Register (AIROC)

  • Led the development of Telstra’s AI Risk Register (AIROC) to track, assess, and mitigate ethical and operational risks associated with AI models enterprise-wide.

  • Designed and implemented the new governance system using Microsoft Power Apps and Power Automate, replacing a legacy manual process with an automated, auditable workflow.

  • Migrated and transformed legacy data to ensure completeness, consistency, and traceability across all AI risk entries.

  • Documented governance processes and conducted handover sessions for new graduates to ensure continuity and compliance.

  • Supported rollout across multiple business units, aligning the platform with Telstra’s Responsible AI Framework and data governance principles.

  • Impact: improved AI governance transparency, reduced reporting effort, and strengthened oversight of ethical and operational risks for all AI initiatives.

  • Tools and technologies: VS Code, Python (Pandas, NumPy, Matplotlib), Power Apps, Power Automate.

Fraud Detection Showcase
QAI Showcase

Biostatistics

HTIN – Machine Learning and AI in Healthcare

  • Developed applied ML skills across healthcare domains, building models for clinical prediction, biomedical signals, imaging, and drug discovery.

  • Applied a broad range of algorithms including regression, clustering, PCA, RNNs, Transformers, and Topic Modelling (LDA).

  • Worked with real-world health datasets for time-series, text, and image data, gaining experience in feature engineering, model interpretation, and evaluation metrics relevant to medical contexts.

  • Tools and frameworks: Python, scikit-learn, PyTorch, DeepReg, and Opacus for privacy-preserving ML.

  • Key outcomes: strengthened understanding of supervised and unsupervised learning, natural language processing for clinical text, and differential privacy in health data, bridging data science with healthcare analytics.

PSI – Principles of Statistical Inference

  • Gained deep statistical foundations critical for data science modelling and uncertainty quantification.

  • Explored likelihood estimation, Bayesian inference, and hypothesis testing to support robust model evaluation and scientific reasoning.

  • Applied maximum likelihood estimation, Wald and score tests, and bootstrap resampling methods to evaluate estimator efficiency and confidence.

  • Developed skills to assess statistical significance, Type I/II errors, and power, informing data-driven decision-making.

  • Key focus areas: estimation theory, Bayesian methods, and inference for model-based analysis.

  • Tools: R and Python for implementing statistical tests and validating model assumptions.

  • Outcome: built strong capability to interpret and validate models statistically, improving rigour in data-driven problem-solving.

DMC – Data Management and Computing

  • Developed strong data engineering and wrangling skills essential for preparing analytical datasets in large-scale environments.

  • Gained hands-on experience in data cleaning, transformation, merging, and relational database design.

  • Applied reproducible programming practices including documentation, scripting, and automated workflows.

  • Created publication-quality visualisations and dashboards for statistical reporting and exploratory analysis.

  • Tools and technologies: R, Python, and Excel; experience with macros, loops, functions, and quality assurance methods.

  • Key outcomes: mastery of ETL processes, data integrity checks, and statistical quality assurance, bridging engineering and analytical workflows.

Machine Learning Project – Mercari Price Suggestion Challenge (UNSW)

  • Developed a predictive pricing model for the Mercari marketplace (1.4M+ listings) to automatically suggest fair and competitive prices for user-listed items, addressing price variability across brands, categories, and item conditions.

  • Cleaned and processed large-scale structured and unstructured data, handling missing values, categorical encoding, and text preprocessing (tokenisation, stop-word removal, stemming).

  • Applied Natural Language Processing (NLP) feature extraction using TF-IDF and Doc2Vec, transforming millions of product descriptions and names into numerical embeddings for model input.

  • Engineered and evaluated multiple regression models (Ridge, Lasso, Support Vector Regression) with extensive hyperparameter tuning via GridSearchCV and cross-validation.

  • Performed dimensionality reduction and log transformation to stabilise variance and improve feature interpretability.

  • Best model: Ridge Regression with TF-IDF features, achieving a Root Mean Squared Logarithmic Error (RMSLE) of 0.477, ranking within the top quartile of the Kaggle competition leaderboard.

  • Demonstrated model interpretability and reproducibility through code documentation, experimentation tracking, and analysis of feature importance.

  • Tools and libraries: Python, scikit-learn, Gensim, Pandas, NumPy, Matplotlib, Google Colab.

  • Impact: delivered a scalable, interpretable machine learning solution that automated product pricing estimation and showcased advanced text analytics for e-commerce optimisation.

Assignment PDF