Abderrahmane Kajeou

Abderrahmane Kajeou

I build fraud-risk models and dashboards, and I check AI model outputs at Amazon.

Data analyst in London. SQL, Python and Power BI on financial and blockchain data, with an MSc dissertation on detecting DeFi fraud on Ethereum.

Email me Download CV GitHub LinkedIn
Data Associate, Amazon AGI
London, since May 2026. Evaluating model outputs for Alexa and Nova in English and French.
97% quality score
On assessed work at Amazon, and Subject Matter Expert on one AI-agent workflow.
MSc Distinction, 85%
University of East London. Dissertation 92%: detecting DeFi fraud on Ethereum.

Projects

DeFi Transaction Risk API

Scores an Ethereum transaction from 0 to 1 and lets the user decide where the alert line sits. Built for my MSc research on leakage-aware risk modelling.

Python, XGBoost, FastAPICode on GitHub

How a transaction becomes a score

  1. TransactionOne Ethereum transfer
  2. FeaturesAbout 46 behavioural signals, built only from the past
  3. XGBoostTrained on earlier blocks
  4. CalibrationIsotonic, so 0.8 means about 80%
  5. ThresholdPolicy set outside the model
  6. ResponseScore, label, latency

Results on the held-out test window

Confusion matrix at threshold 0.287. Darker means more transactions.
Model said low riskModel said high risk
Actually legitimate30,209correctly passed995false alarms
Actually fraud3,724missed23,751caught

58,679 test transactions. Counts are read from the confusion matrix in the repository.

Precisionof flagged transactions, how many were fraud
96.0%
Recallof all fraud, how much was flagged
86.4%
False alarm ratelegitimate transactions flagged
3.2%
ROC-AUC and PR-AUCoffline, as reported in the repository
0.95 and 0.94
Screenshot of the DeFi Fraud Detector interface showing a transaction form, a risk probability dial, the JSON response and a log of recent predictions.
The monitoring interface. Enter a transaction, test a different threshold without retraining, and read the response and latency from a prediction log.
What this does not show
  • All results come from offline experiments on a public phishing dataset. There is no live blockchain connection.
  • The model uses tabular behaviour features only, with no transaction graph or contract analysis.
  • Ranking quality drops under strict time-based splits, so I present it as a triage prototype and not a fraud verdict.

Invoice Proof Registry

Stops invoice tampering and duplicate claims by storing a hash of each invoice on-chain, with role-based approval and an audit trail.

Solidity, Hardhat, ethers.js, OpenZeppelinCode on GitHub

Invoice lifecycle and who can act

  1. CreatedIssuer registers the hash
  2. Approved or rejectedApprover decides
  3. PaidFinance records payment
Screenshot of the registry interface showing the workflow buttons and an output panel reporting match true.
Verification recomputes the document hash and returns match: true only if it equals the registered hash.

What it does

Only the Keccak256 hash and metadata go on-chain, never the document. Smart-contract roles gate each action, every step emits an event, and a timeline view shows the history of any invoice. A CSV and SQL pipeline turns on-chain data into audit queries.

A local proof of concept on a Hardhat test chain. It is not audited and not for production use.

Fraud Risk Monitoring Dashboard

Power BI dashboard over 1.3M+ transactions worth £91M. It tracks fraud trends, flags abnormal spikes and high-risk periods, and filters suspicious accounts and transaction chains for investigation.

Power BI, SQL

Experience

Data Associate, Amazon AGISME on one workflow

Amazon, London

  • Evaluate and annotate multimodal data (text, image, audio, video) in structured ML evaluation pipelines that support Alexa and Nova.
  • Assess AI-generated outputs for accuracy, relevance and linguistic quality in English and French against defined criteria.
  • Find systematic errors in model outputs and annotation patterns, and feed quality signals back into LLM training data pipelines.
  • Hold a 97% quality score and act as Subject Matter Expert for one of the workflows supporting an AI agent.
  • Worked on an L4 workflow evaluating animation-related model outputs, extending evaluation into visual and motion content.

Internal Control Analyst Intern

National Office of Electricity and Drinking Water (ONEE), Oujda

  • Built Power BI dashboards tracking operational KPIs across 10+ reporting areas, cutting manual Excel reporting time for the internal control team.
  • Analysed financial transaction data to find spending patterns and inefficiencies for internal audit reporting.

Financial Delegate Intern

Faculty of Economic, Social and Legal Sciences, Oujda

  • Conducted 3 market analyses that identified financial opportunities and informed budget and cost decisions.
  • Managed 500+ financial records on the AD@E platform, improving accuracy and audit readiness.

Finance Intern

Presidency of Mohamed 1 University, Oujda

  • Supported financial reporting for a department with a $500K budget.
  • Coordinated a $200K departmental budget, improving tracking and resource allocation.

Blockchain and crypto market research

Self-directed

  • Researched 50+ blockchain and DeFi projects: tokenomics, risk factors and on-chain transaction behaviour.
  • Analysed Bitcoin market cycles around halving events in Python and Excel, with time-series models of price behaviour before and after.

Skills

Data analysis
SQL (joins, CTEs, window functions, subqueries), Python (Pandas, NumPy, scikit-learn, Matplotlib, Seaborn), Excel (Power Query, pivot tables), Power BI (DAX, data modelling, interactive dashboards)
Machine learning and risk
Classification, anomaly and fraud detection, XGBoost, probability calibration, risk scoring systems, probability foundations for data science
AI data and evaluation
Multimodal annotation (text, image, audio, video), model-output quality assessment, structured evaluation workflows in English and French
Tools and blockchain
Git, Jupyter, REST APIs, FastAPI, Solidity, ethers.js, DeFi and on-chain transaction analysis
Languages
Arabic (native), French (upper-intermediate), English (upper-intermediate)

Education and certifications

MSc Blockchain and Financial Technology, Distinction (85%)

University of East London

Dissertation, 92%: Detecting DeFi Fraud on Ethereum, an AI-driven approach to smart-contract risk mitigation.

Professional Bachelor's in IT Business Management, ranked 3rd in cohort

Higher School of Technology, Oujda

Final year project, 16/20: performance monitoring dashboards for decision-making.

DUT in Finance, Accounting and Taxation

Higher School of Technology, Oujda

Probability Foundations for Data Science and AI

University of Colorado Boulder, on Coursera

Excel Skills for Business, Macquarie University

Microsoft Power BI Data Analyst, in progress

Contact

I am looking for a Data Analyst, BI Analyst or AI Data and Quality role in London. Email is the quickest way to reach me.

abdoukaj8@gmail.com
Download CV LinkedIn GitHub