# Owen Cheung — AI Engineer AI Engineer at LexisNexis Risk Solutions, building human-governed workflows for fraud-model optimisation\. Final-year Computer Science & AI student at the University of Bath\. I'm Owen, an AI Engineer at LexisNexis Risk Solutions and a final-year Computer Science & AI student at the University of Bath\. I stayed on as a contractor after my industrial placement to develop a self-initiated prototype into a funded internal pilot\. I enjoy the freedom to investigate a problem, explore emerging methods and put promising ideas to the test\. My work connects representation-learning research with practical AI engineering: from graph and contrastive methods to recoverable, human-governed systems\. Away from the screen I snowboard with Bath Snowsports, pull a lot of espresso, and build mechanical keyboards — which is how I ended up typing at 193 wpm (and designing TypeForge)\. Contact: oc608@bath\.ac\.uk LinkedIn: GitHub: ## Experience ### AI Engineer (Agentic AI) — LexisNexis Risk Solutions Aug 2026 — present · Contract Architected a solution designed for organisation-wide use, built an end-to-end demo, and presented the concept, value and scaling approach to senior stakeholders\. Now developing the funded internal pilot toward production\. ### Data Scientist — LexisNexis Risk Solutions Jul 2025 — Aug 2026 · Industrial placement Large-scale financial data analysis, two fraud-model optimisations and self-directed representation-learning research\. Gained experience in GPU compute and batch training, and initiated an agentic working demo\. ### Full-Stack Developer — Dish'D Jan 2024 — Feb 2025 Built a social cooking app end-to-end with Django \+ Flutter; led technical decisions\. ### System Tester — Kinetix Jul 2024 — Aug 2024 QA for Census & Statistics Department systems — validated functionality against spec\. ### AI Project Intern — DXC Technology Jul 2022 — Aug 2022 Drafted technical proposals for enterprise bids (AI-assisted mapping, OpenBIM\.AI)\. ## Education University of Bath — BSc (Hons) Computer Science & Artificial Intelligence, with placement · Sept 2023 — 2027 (expected) · Expected 2:1 Wellington College — A-Levels — Maths, Further Maths, Physics, Computer Science · Aug 2021 — Aug 2023 · ## Skills - AI / ML: PyTorch, Contrastive learning, Hyperbolic embeddings, Agentic systems, Deep learning, Multimodal contrastive learning (study) - Languages: Python, TypeScript, Haskell, Dart, SQL - Frameworks: React / Node, Django, Flutter, Next\.js - Data: SQL, Snowflake, LightGBM, Statistical modelling, Data analysis - Compute: GPU compute, Batch training - Cloud / AI: Azure OpenAI, Azure Foundry and AWS S3 (workshop experience) ## Capabilities ### ML Research PyTorch · Multimodal contrastive learning (study) Following a question beyond the obvious answer\. My research explores how models learn useful representations of complex behaviour\. - Researched self-supervised representations using graph embeddings, contrastive learning and non-Euclidean geometry\. - Used the autonomy of my role to investigate emerging methods and experiment with their application to practical problems\. - Multimodal contrastive learning is an area of study, not a claim of an implemented or deployed multimodal system\. ### AI Systems Python · Agent orchestration · Human review Taking an idea from architecture to an end-to-end demonstration, and making its value clear to the people deciding what comes next\. - Designed an agentic solution for organisation-wide use, with persistent state, audit trails and mandatory human-review gates\. - Built an end-to-end working demonstration\. Presented the core concept, value and scaling approach to senior stakeholders\. - Retained to develop a funded internal pilot\. Intended scope is not a claim of organisation-wide deployment\. ### Data & Modelling SQL · GPU compute · Batch training Working with large-scale financial data, then connecting analysis, modelling and evaluation to a concrete problem\. - Analysed financial transaction data with Python, SQL and Snowflake, and completed two fraud-model optimisations during placement\. - Applied feature engineering, hyperparameter tuning and performance evaluation\. - Gained experience in GPU compute and batch training; no public throughput or model-performance benchmark is claimed\. ## Interests - Snowboarding: Carving and freestyle with Bath Snowsports — away from the screen\. - Specialty coffee: Chasing brew methods and flavour\. - Mechanical keyboards: Built several from scratch; reached 193 wpm\. ## Projects # LexisNexis — Applied AI Internal pilot · toward production · Jul 2025 — present Role: AI Engineer (Agentic AI) · previously Data Scientist Stack: Python, SQLite, DVC, Papermill, LightGBM I architected a fraud-model optimisation solution designed for organisation-wide use and built an end-to-end working demonstration\. I presented its core concept, business value and approach to scaling to senior stakeholders\. After my placement, I was retained as an AI engineering contractor to develop the funded internal pilot toward production\. ## Context Fraud-model optimisation combines analysis, modelling and decisions that need accountable human review\. ## My contribution Architected a solution designed for organisation-wide use, built an end-to-end demonstration, and presented its concept, value and scaling approach to senior stakeholders\. Retained after placement to develop the funded internal pilot\. ## Approach - Custom Python orchestration and reusable Copilot skills coordinate work with persistent SQLite state\. - Audit trails and mandatory review gates keep human judgement in the workflow\. - DVC and Papermill make notebook-based work recoverable; LightGBM discovers candidate rules, with constrained logistic regression for policy weighting\. ## Highlights - Custom Python orchestration, reusable Copilot skills, persistent SQLite state, audit trails and mandatory human review gates\. - Recoverable DVC/Papermill workflows with LightGBM candidate-rule discovery and constrained logistic-regression policy weighting\. - Built an end-to-end demonstration and presented the core concept, business value and scaling approach to senior stakeholders\. ## Evidence - End-to-end working demo delivered by the end of the placement\. Presented the core concept, value and scaling approach to senior stakeholders\. - Retained as an AI engineering contractor in August 2026 to develop a capital-funded internal pilot\. ## Limitations - Employer-confidential work: no source code, transaction data or internal evaluation results are published\. - This describes a pilot building toward production, not a launched production platform\. - Organisation-wide use describes the intended architectural scope, not a claim of organisation-wide deployment or tested production scale\. - Descriptions are owner-supplied; no independently verified performance uplift is claimed\. Source: Owner-supplied professional and project descriptions; earlier non-conflicting portfolio material retained\. Not independent verification\. # LexisNexis — ML Industrial placement · Jul 2025 — Aug 2026 Role: Data Scientist · Industrial placement Stack: Python, SQL, Snowflake, Feature engineering, Representation learning During my industrial placement at LexisNexis Risk Solutions, I completed two fraud-model optimisations and analysed large-scale financial data using Python, SQL and Snowflake\. Alongside this applied work, I researched self-supervised fraud representations and gained practical experience in GPU compute and batch training\. ## Context Fraud modelling involves understanding complex financial behaviour, developing useful features and evaluating changes carefully\. ## My contribution Completed two fraud-model optimisations and analysed large-scale financial data\. Researched self-supervised representations alongside practical modelling work\. ## Approach - Used Python, SQL and Snowflake to analyse large-scale financial data\. - Worked across feature engineering, model tuning and evaluation for fraud-model optimisation\. - Investigated graph embeddings, contrastive learning and non-Euclidean geometry as approaches to self-supervised fraud representations\. - Gained practical experience in GPU compute and batch training\. ## Highlights - Completed two fraud-model optimisations, working across feature engineering, tuning and evaluation\. - Analysed large-scale financial data using Python, SQL and Snowflake\. - Researched graph embeddings, contrastive learning and non-Euclidean geometry for self-supervised fraud representations\. - Gained practical experience in GPU compute and batch training\. ## Evidence - Two fraud-model optimisations completed during the industrial placement, July 2025 to August 2026\. - Research into self-supervised fraud representations accompanied applied data-science work\. ## Limitations - Employer-confidential work: no transaction data, source code or internal evaluation results are published\. - Research directions are not claims of deployed models or measured performance uplift\. - Descriptions are owner-supplied; no independently verified results are claimed\. Source: Owner-supplied professional and project descriptions; earlier non-conflicting portfolio material retained\. Not independent verification\. # TypeForge In development · Jul 2025 — present Role: Solo developer Stack: React, Node\.js, TypeScript, Keystroke analytics I am developing a typing application that tracks cadence, keystrokes and accuracy, with AI-driven drills designed and implemented to adapt to user weaknesses\. It comes from building mechanical keyboards and reaching 193 wpm\. A public demo is available; the product remains in development, with no public evaluation or full product release claimed\. Live demo: ## Context Typing practice can focus on a headline speed rather than the particular weaknesses a person needs to practise\. ## My contribution Designing and building a React/Node\.js typing application, including implemented AI-driven adaptive drills\. ## Approach - Developing real-time analytics for cadence, keystrokes and accuracy\. - Adaptive drills target user weaknesses\. - Personal typing and keyboard-building experience informs the product direction\. ## Highlights - Designed and implemented adaptive AI-driven drills\. - Developing real-time cadence, keystroke and accuracy analytics\. - A personal exploration of more targeted typing practice\. ## Evidence - Adaptive AI-driven drills designed and implemented, as described in my current project record\. ## Limitations - A public demo is available; the product is still in development, not a full product release\. - No measured learning uplift, timing precision or model benchmark is published\. Source: Owner-supplied professional and project descriptions; earlier non-conflicting portfolio material retained\. Not independent verification\.