Machine learning projects

Machine learning case studies by Kanishka Ranaweera, PhD, ML engineer in Melbourne: federated and private ML, computer vision, time-series, probabilistic modelling and engineering R&D.

Medical Imaging

• Medical CT Segmentation: This work is proprietary and cannot be disclosed.
• Medical X-Ray Analysis: This work is proprietary and cannot be disclosed.
• MRI Studies of Learning in Neurological Disorders: Machine learning on brain MRI to study how neurological disorders relate to learning abilities.

AI Products & Platforms

• AI Receptionist for Car Dealerships: An AI receptionist that books test drives and answers customer questions using retrieval-augmented generation.
• AI Physiotherapist for Guided Rehabilitation: An AI physiotherapist at Biokin that guides patients through their exercises and tracks how accurately each activity is performed.
• Practice Management System for Accounting Firms: A practice management system that brings an accounting firm's clients and work into one platform.
• AI Bank Reconciliation for Accountants: An AI-assisted bank reconciliation system that matches bank transactions against accounting records for accountants.
• AI BAS Calculation System for Accountants: An AI-powered system that helps accountants calculate Business Activity Statement (BAS) figures.
• Support Ticketing System for Software Products: A support ticketing system built to be used inside software products, so users can raise and track issues.
• One Focus Finance: Home Loan Calculators: Five home-loan calculators (borrowing power, repayments, offset, stamp duty and deposit) inside a FinTech platform for Melbourne mortgage brokers.
• Website Design and Development: Designed and built websites for Merit Grove and Thurstan College OBU Australia.
• AusCey TV: News and Events Platform: A news and events platform connecting Australia and Sri Lanka through news, programs, live TV and community stories.
• LagnaVerse: AI Vedic Astrologer: Instant Vedic birth charts, a conversational AI astrologer and traditional Porondam matching, in four languages.

Regulated AI & Compliance

• NextPhase: Post-Market Surveillance for ML Models: Open-source post-deployment monitoring that wraps any model and tracks drift, data quality and performance once it is live.
• A 21 CFR Part 11 and ISO 13485 Compliant Electronic Document Management System: A full EDMS for controlled documents, with electronic signatures, multi-stage approvals and a complete audit trail.
• An LLM-Assisted QMS That Auto-Scores Regulatory Readiness: Scores documentation against individual regulatory criteria and uses RAG to map evidence and draft what is missing.

Federated & Private ML

• IRT-FedLoRA: Reliability-Aware Federated Fine-Tuning: A probabilistic IRT model estimates client reliability over time and uses it to weight federated LoRA updates.
• ClusterGuardFL: Secure Cluster-Weighted Client Aggregation: Client updates are clustered and weighted using behavioural similarity and reconciliation confidence rather than treated equally.
• AR-FL: Anomaly-Resilient Federated Learning: Adaptive anomaly detection and weighted aggregation reduce the impact of faulty or adversarial client updates.
• DP-FedMeta: Differentially Private Federated Meta-Learning: Adaptive clipping improves the privacy-utility trade-off in personalised federated meta-learning.
• Utility-Enhanced Differentially Private Federated Learning: Haar-wavelet transformation and a modified noise-injection strategy improve convergence under differential privacy.
• Multi-Objective Privacy-Utility Optimisation in Federated Learning: A multi-objective optimiser adapts gradient clipping to balance model accuracy and privacy protection.
• Meta-Clip: Privacy-Preserving Few-Shot Learning: Adaptive clipping helps few-shot meta-learning retain generalisation while protecting sensitive training data.

Time-Series & Sensors

• Probabilistic Deep Ensembles for FRDA Severity Assessment: Complementary time-series networks produce both mFARS predictions and interpretable uncertainty intervals.
• Cross-Device Domain Adaptation for FRDA Assessment: Adversarial domain adaptation transfers knowledge between instrumented cup and spoon data to improve severity prediction.
• AIM-M: Multiview IMU Assessment of Friedreich Ataxia: A multiview model combines cup, spoon and pendant sensors into a single objective severity score.
• DTW-Based Outlier Removal for Reliable FRDA Measurement: Dynamic Time Warping filters non-representative sensor trials before feature extraction and severity modelling.
• Learning Using Privileged Information for FRDA Assessment: A LUPI framework uses genetic and disease-history information during training while requiring only sensor data at deployment.
• Forecasting Loss of Ambulation in Friedreich Ataxia: IMU, genetic and clinical data are combined with accelerated failure-time models to forecast transitions in ambulatory status.
• Reliable FRDA Assessment with Isolation Forest Anomaly Detection: Isolation Forest filtering identifies irregular AIM-C trials before they distort disease-severity predictions.
• Objective FRDA Assessment in Children: Device-specific correction models separate developmental movement patterns from disease-related impairment in children.
• Dual-Relationship Modeling of Friedreich Ataxia: A dual-objective model learns both current disease severity and biologically plausible progression over time.
• Adaptive LSTM-RNN Dynamic Equivalents for Power Systems: An adaptive LSTM-RNN reduces power-system model complexity while maintaining accuracy across changing operating points.

Computer Vision

• Pneumonia Detection with Vision Transformers: Vision Transformers capture long-range spatial relationships in chest X-rays for high-accuracy pneumonia classification.
• Few-Shot Medical Diagnosis with Model-Agnostic Meta-Learning: A 20-shot MAML strategy rapidly adapts to new diagnostic tasks when labelled medical images are scarce.
• Robot Vision for Search-and-Rescue Facial Recognition: A robot-vision pipeline explores face localisation, preprocessing and recognition for hazardous search-and-rescue environments.

Probabilistic Modelling

• FAIR: Interpretable Latent Severity Scoring for Neurological Disorders: A Bayesian Item Response Theory framework for creating an interpretable latent disease-severity score from multiple clinical scales.
• FA-DPS: A Composite Severity Score for Friedreich Ataxia: A biologically grounded disease-progression score provides a more suitable machine-learning target than a single subjective clinical scale.

Engineering R&D

• Aquacell PPE Research & Prototype Development: A concept-to-prototype R&D project for a novel personal protective facemask intended to reduce transmission of COVID-19 and similar viruses.