ACADEMIC PROJECTS / 2026
Neural networks, deep learning & cloud engineering.
Selected coursework from CQUniversity. These entries document experimental methods, evaluation results and specific implementation scope. The coursework source repositories are now public. Explore the individual evidence-backed case studies below or follow their GitHub links. Raw datasets and generated deep-learning models may still be excluded from those source repositories.
01 / NATURAL LANGUAGE PROCESSING
BBC News Classification
Compared a TF-IDF feature pipeline and a Word2Vec-based representation, each feeding a multilayer perceptron, to classify five news topics: Business, Entertainment, Politics, Sport and Technology.
- Held-out test
- 426 articles
- TF-IDF + MLP
- 97.18% test accuracy
- Word2Vec + MLP
- 96.01% test accuracy
Engineering approach
- Validated the dataset and removed exact duplicates before a reproducible stratified split into training, validation and test subsets.
- Fitted text representations on training data only to reduce information leakage.
- Compared model performance using held-out predictions, confusion matrices, precision, recall and F1.
The 97.18% figure is the recorded test-set accuracy on this BBC News coursework dataset, not a claim of general performance on arbitrary news.
02 / NEURAL NETWORK EXPERIMENTS
DNN Regression, Classification & Data Analysis
Developed experiments for California and Brisbane housing-price regression and olive-oil classification, with supporting PCA, K-means clustering, feature analysis and visualisations.
- California regression
- Test R² ≈ 0.783
- Brisbane regression
- Test R² ≈ 0.970
- Olive-oil classification
- 99.2% recorded accuracy
Engineering approach
- Compared hidden-layer depths, neurons, batch sizes, learning rates, optimisers and epoch counts using logged experiments.
- Combined supervised neural networks with unsupervised analysis to explore data distribution and feature relationships.
- Saved training curves, prediction plots, summary CSVs and a final two-class olive-oil classification report.
These are results recorded in the assessment repository's experiment outputs for the specified datasets and splits; they are not production model benchmarks.
03 / CONVOLUTIONAL NEURAL NETWORKS
CNN Age & Gender Estimation
Implemented a shared convolutional feature extractor with separate age-regression and gender-classification heads using the UTKFace image dataset, comparing the original model with a data-augmented training approach.
- Augmented model
- 6.871-year validation age MAE
- Augmented model
- 88.76% validation gender accuracy
- Original model
- 89.35% validation gender accuracy
Engineering approach
- Normalised 128 × 128 RGB images and compared baseline and augmented CNN configurations.
- Applied horizontal flips, rotation, zoom and contrast variation during augmented training.
- Recorded learning curves and tuning experiments; noted the trade-off between lower age error and gender classification accuracy.
Validation results are experimental model measurements. Age and gender inferred from images can be inaccurate or biased and are not used for real-world identity or personal assessment.
04 / TIME-SERIES DEEP LEARNING
LSTM Mining Stock Forecasting
Implemented recursive LSTM price forecasting for Rio Tinto (RIO), BHP and Fortescue (FMG), using 5-day and 10-day historical lookback settings and comparing predictions against a simple last-observed-price baseline.
- Experiments
- 36 tuning configurations
- Forecast horizon
- 50 held-out trading days
- Baseline comparison
- LSTM better for RIO and BHP, not FMG
Engineering approach
- Used chronological train/test separation and training-only scaling during parameter selection.
- Evaluated stock-specific RMSE against a last-price baseline rather than reporting model predictions alone.
- Documented the important negative result: FMG's baseline RMSE of A$0.832 was lower than the best LSTM RMSE of A$1.569.
This was a retrospective academic experiment using historical price series, not a trading strategy or a guarantee of future market performance.
05 / AWS CLOUD COMPUTING
Cloud-Connected Temperature Monitoring
A collaborative AWS smart-monitoring design in which a Python simulator publishes test sensor readings through certificate-authenticated MQTT into AWS IoT Core. The broader team architecture integrates event processing, storage, notification and monitoring services.
- My documented workstream
- IoT ingestion and simulator
- Transport
- MQTT over TLS with X.509
- Cloud environment
- AWS Sydney region
Scope and attribution
- My documented ingestion work includes the Python simulator, AWS IoT Thing and certificate/policy setup, and publishing controlled normal, high and low readings.
- The team design uses AWS IoT Core, Lambda, DynamoDB, SNS, CloudWatch and IAM for secure processing, persistence, abnormal-reading notifications and observability.
- The repository's README records the ingestion component as completed but describes Lambda/storage/alerts and full end-to-end integration as remaining work; I am not claiming those broader services were independently completed by me.
Public source: AWS temperature monitoring group repository (opens in a new tab) ↗.
AWS account details, IoT certificates and private keys are intentionally excluded from this public portfolio.
06 / DATA EXPLORATION
Brisbane House Price Analysis
A separate 110-property descriptive-analysis project comparing price categories and demand across five Brisbane suburbs.