Machine Learning PhD Candidate · Western University

Rashinda Wijethunga

PhD in Machine Learning, expected August 2026. For two years I worked with Neelands Group as an Applied Machine Learning Research Intern, deploying two ML systems for early slow-leak detection in supermarket refrigeration systems; before that, a Systems Engineer in Sri Lanka. I've built two agentic Generative AI projects and research GNN-based time-series anomaly detection and explainable, trustworthy AI — with 6 peer-reviewed publications, including top-tier Q1 journals.

Peer-reviewed publications

Industry ML deployments

Generative & Agentic AI projects

$50K

In research grants

Years building ML

Selected work

Featured Work

Featured Generative & Agentic AI · Live demo

BuildRight AI — Guardrailed, Agentic RAG Commerce Assistant

A full-stack hardware-store assistant: an agentic Claude tool-use loop over a 10K+ SKU catalog, hybrid (lexical + vector + visual) RAG with reciprocal-rank fusion and re-ranking, and a deterministic guardrail that validates every price and citation against retrieved data — so a fabricated price can't render. A Haiku→Sonnet router keeps it cheap; it ships as one Docker container.

Hybrid RAG CLIP visual search Deterministic guardrails Haiku→Sonnet routing

Read the case study → Live demo ↗

Agentic AI

LLM-Powered Multi-Agent Chatbot for Enterprise User Provisioning

A five-agent system (Google ADK, Gemini) automating onboarding/offboarding via a supervisor delegating to policy, identity, and approval agents — with human-in-the-loop gates and a custom MCP server over live Entra ID.

Google ADK Gemini MCP A2A

Explore →

Industry Deployment · Mitacs

Supermarket Refrigerant Leak Detection

ML leak-detection deployed across supermarket refrigeration systems — cut detection latency from months to days across 170+ subsystems per site. Shipped on Azure Functions with CI/CD, automated retraining, and drift monitoring.

XGBoost PySpark Azure MLOps

Explore →

GNN Research

DualSTGF — Dual Spectral-Temporal Graph Fusion

Dual-view spectral-temporal graph fusion with adaptive edge learning and divergence scoring for incipient fault detection. Submitted to Information Fusion (IF 15.5).

PyTorch PyG Spectral GNN

Explore →

GNN Research · Published

Dual-Graph GNN for Structural Damage Detection

Dual graph representations for structural health monitoring, with a 1D-CNN compression module that cut training time 24×. Published in Engineering Structures (Q1, IF 6.4, 2025).

PyTorch PyG SHM

Explore →

View all projects →

What I do

Capabilities

Generative & Agentic AI

Guardrailed RAG, multi-agent tool-use loops, LLM cost routing, memory, and evaluation harnesses — built on raw SDKs (Anthropic, Google ADK, MCP) for full control over grounding and safety.

ML in Production

From paper to deployment: FastAPI/Docker services, CI/CD, drift monitoring, and per-turn observability. Two ML systems deployed to production for supermarket refrigeration via Mitacs Accelerate.

Trustworthy ML & GNNs

Spatio-temporal graph neural networks and counterfactual explainability for anomaly detection — the focus of my PhD and 6 peer-reviewed papers, including top-tier Q1 journals.

Background

Experience

From a PhD on trustworthy graph neural networks to two industry systems shipped through Mitacs Accelerate, plus independent Generative & Agentic AI work.

Applied Machine Learning Research Intern (Phase II)

Jan – Dec 2025

Neelands Group · Remote (Burlington, ON)

Built and deployed an end-to-end anomaly-detection system for CO₂ supermarket refrigeration (170+ subsystems per site). Processed ~210M rows of telemetry with PySpark; trained per-site XGBoost with residual-based detection in a state-machine architecture, hitting an 87.5% detection rate. Shipped on Azure Functions with CI/CD, automated retraining, and drift monitoring.

Applied Machine Learning Research Intern (Phase I)

May 2022 – Apr 2023

Neelands Group · Remote (Burlington, ON)

Designed a sensor-independent leak-detection framework for HFC refrigeration using only standard telemetry. Physics-informed features with SHAP-driven selection; a CatBoost model with dynamic thresholding cut detection latency from months to ~5 days at 100% precision (F1 0.92). Success led to a second engagement at the same client.

PhD Researcher

Sep 2021 – Aug 2026 (expected)

Western University · London, ON

Dissertation on trustworthy anomaly detection with graph neural networks and counterfactual explanations. Built an unsupervised temporal GNN that beat 9 baselines by up to 7% AUC; a 1D-CNN compression module cutting training time 24× (296 h → 12 h); and a robust tree-ensemble explainability method ranked #1 in 158/248 perturbation scenarios.

Systems Engineer

Mar 2020 – Aug 2021

Millennium IT ESP · Colombo, Sri Lanka

Engineered the migration of legacy data-center infrastructure to Cisco's application-centric (ACI) architecture for SLT, Sri Lanka's largest telecom provider.

Education

Ph.D., Electrical & Computer Engineering

Western University · 2021 – 2026 (expected)

Trustworthy Anomaly Detection using Graph Neural Networks & Counterfactual Explanations

B.Sc. Eng. (Hons), Electronic & Telecommunication

University of Moratuwa · 2015 – 2020

Awards & Certifications

Mitacs Accelerate Scholarships — $30K (2025) + $20K (2022–23)
IEEE ComSoc Student Competition — Honorary Mention, World Top 15 (2019)
5-Day AI Agents Intensive — Google & Kaggle

Read the full bio →

Toolbox

Skills & Tools

Tools & frameworks

Python

PyTorch

scikit-learn

TypeScript

React

FastAPI

Docker

PostgreSQL

Gemini

Hugging Face

PySpark

Stripe

Methods & concepts

Hybrid retrieval · RRF · re-ranking Grounding & guardrails RAG evaluation / faithfulness Fine-tuning Multi-agent systems Tool-use / function calling Memory & human-in-the-loop Model Context Protocol (MCP) Agent-to-Agent (A2A) Graph Neural Networks Counterfactual explainability Anomaly detection Synthetic data generation

Research