Evan Harley

Full Stack Data Scientist | ML/AI Strategy | Data Engineering

10+ years building end-to-end data solutions — from lakehouse architecture and production pipelines to predictive models and AI adoption strategy.

About Me

Over 10 years in data science, I’ve worked across the full analytics lifecycle — pipeline architecture, data engineering, predictive modelling, NLP, computer vision, and ML/AI strategy. I’m equally comfortable presenting an AI roadmap to a Deputy Minister and writing the production pipeline that underpins it.

In the BC Public Service I’ve applied this breadth to high-stakes domains: designing a ministry-wide Data Lakehouse, authoring the AI Adoption Strategy, applying survivorship analysis and event-duration prediction to provincial roadway incident data, and advising senior leadership on agentic AI use cases. The work is meaningful, the problems are genuinely hard, and the scale is real.

Outside the office I’m usually satisfying curiosity in other directions — training in classical Japanese martial arts, running a homelab Kubernetes cluster, or diving into medieval history. I’m a firm believer that the best ideas come from a broad range of experiences, and I’m always looking for the next interesting problem to solve.

Your Image Here

Featured Projects

Nyström-OCRKM: Scalable Anomaly Detection

Implemented a One-Class Restricted Kernel Machine and scaled it to millions of ride-hail trip records via Nyström approximation (O(N²) → O(Nm)), benchmarked against Isolation Forest and SGD-OCSVM. It surfaced four classes of data-quality defect that pass every rule-based check — each independently corroborated — while honestly reporting where the simpler baselines won. A graduate research project written up as a conference-style paper.

PythonKernel MethodsNyström ApproximationAnomaly Detectionscikit-learn

Public-data (NYC TLC) rerun in progress — code link to follow.

Base45: Hypertrophy Training App + MLOps Research Layer

A multi-user web app that automates progressive-overload programming for hypertrophy — turning each session’s logged sets into the next session’s prescribed sets, load, and reps. Built end to end on a MotherDuck medallion lakehouse with an MLflow research loop on top; that research surfaced a "controller-censoring" effect — a well-tuned training controller erases the very variance naive models chase — reframing the ML work toward honest, diagnostic signal.

SvelteKitTypeScriptMotherDuckMLflowKestraApplied ML

Skills & Technologies

Languages & Databases

Languages

Python (10+ yrs)R (10+ yrs)SQLTypeScript / JavaScriptC#C++Dart / Flutter

Databases

PostgreSQLSQL ServerMongoDBDuckDB
ML / AI

Methods

Supervised & Unsupervised LearningNLPComputer VisionTime Series ForecastingSurvival Analysis

Modern AI

LLMsAgentic AIRAGFine-Tuning
Data Engineering & MLOps

Engineering

Data Lakehouse ArchitecturePipeline DesignETL / ELTDuckDB / MotherDuckEstuary Flow

Deployment

DockerKubernetesSvelteStreamlitGitHub Actions
BI, Visualization & Strategy

BI & Visualization

Power BITableauPlotlyMatplotlibggplot2

Strategy & Leadership

AI Adoption StrategyData StrategyExecutive CommunicationWorking Group Leadership

Let's Connect

I'm always open to discussing new opportunities and interesting projects. Feel free to reach out or explore my work further.