Fraud investigator turned analyst.
Same instincts, sharper tools.
Two years of front-line fraud investigation at HSBC and Monzo, now completing an MSc in Business Analytics at UCL. I have adjudicated thousands of outputs from production fraud-detection models, and I am now building and evaluating models myself.
About
I started on the front line of UK banking, spotting fraud at account opening for HSBC, then investigating APP scams there, then adjudicating model-flagged fraud alerts at Monzo with my own £10,000 reimbursement authority. Along the way I kept finding problems that sat above my job description: a payment feature that could move £25,000 without warning the customer (taken down after I escalated it), false-positive patterns worth feeding back into detection thresholds, an incident where push notifications contradicted what was happening at the tills.
That pattern, noticing where a system's outputs do not match reality, is what took me to UCL. The MSc is the technical depth behind the instinct: building classifiers, evaluating language models, designing guardrails for AI systems. My dissertation, in partnership with Lifemote, clusters WiFi telemetry from over half a million home gateways into failure archetypes ISPs can act on proactively. I speak English and Arabic fluently, plus conversational German. The direction of travel: model risk, model validation, and second-line financial crime analytics, where investigative discipline and statistical literacy meet.
Experience
UCL-mediated dissertation partnership with Lifemote, a WiFi analytics scaleup serving ISP networks across roughly 6 million UK and European households. Built a two-stage unsupervised clustering pipeline in Python (PCA, K-Means against a GMM comparator) over a weekly telemetry snapshot of 525,983 home gateways, identifying failure archetypes and validating them internally, against a held-out quality-of-experience label, and across consecutive weeks.
Adjudicated model-flagged fraud alerts across APP scams, card fraud and account takeover, with reimbursement authority up to £10,000 and a caseload of roughly 120 to 160 cases a week. Maintained a 95%+ QA score, validated by the QA team and line manager.
Identified recurring false-positive patterns and escalated them to the fraud detection function for threshold and feature recalibration. Raised the Easter 2025 M&S till-outage notification discrepancy as a formal incident. Caught first-party fraud running through the chat-bot refund flow and closed repeat offenders' accounts under my own authority.
Investigated authorised push payment scam cases, applied the CRM Code to assess reimbursement eligibility, and produced audit-ready documentation for every outcome. 100% QA pass rate across closed cases, externally scored.
Escalated a flaw in HSBC's new Pay by Bank App feature, which could trigger transfers of up to £25,000 with no prior customer notification. The feature was subsequently taken down.
Front-line banking role where the fraud-detection instinct first became visible: identified fraud risks at onboarding, applied customer due diligence principles, and was invited into product-design meetings on HSBC's new app to advise on the strength of customer-facing fraud warnings.
Projects
Formal graded coursework from the UCL MSc Business Analytics programme (2025/26). Individual projects are entirely my own work. Group projects state the team size and my specific contribution.
Two-stage unsupervised clustering over a 525,983-gateway telemetry snapshot: PCA, K-Means vs GMM, validated internally, against a held-out QoE label, and across consecutive weeks.
F1-optimised classification pipeline on imbalanced booking data: explicit sparsity handling, leakage controls, and a calibrated decision threshold. The same mechanics that sit inside fraud detection models.
Streamlit research assistant with a separate LLM-as-judge evaluation harness, semantic input validation, and prompt guardrails written to stop the model fabricating figures.
Four systems built from first principles in one notebook: sentiment classifiers, a PyTorch MLP against LLM zero-shot baselines, TF-IDF retrieval with NER, and n-gram language models.
Mobile MVP combining a deterministic burnout-guard scoring layer with a Claude-powered tool-calling agent over live event APIs. Every suggested venue verified against a live API before display.
Controlled study of Claude Code, Codex and Antigravity across a five-task pipeline. Headline finding: the agent with the perfect raw score needed the most supervision, and pricing interventions in reordered the ranking.
Do sentiment classifiers hold up across product categories? Category-stratified evaluation of classical and transformer models, plus topic modelling on the reviews they got wrong.
Additional coursework
Willingness-to-pay and net-utility consumer-choice modelling in Excel to recommend a revenue-maximising personal-training pricing strategy across flat, peak-load and segmented options for a 10,000-member health-club case.
Excel inventory models quantifying drop-shipment, shipping-versus-holding and assembly-cost trade-offs under stochastic demand, producing SKU-level distribution recommendations.
Contact
Based in London, open to UK remote. Interested in model risk, model validation and second-line financial crime analytics roles.