Data scientist for payments, fraud and risk analytics in New York
I am Syed Ahmed Basharat Ali, a data scientist and analytics engineer in New York City. Most of my payments experience comes from Geidea, Saudi Arabia's largest payments processor, where I was a Data Scientist from Aug 2023 to Jan 2025. I was promoted from Business Analyst within 6 months and was a member of the executive strategy team.
My work there covered fraud and anomaly detection, merchant risk and churn, settlement data and collections. Since then I have published public projects that price fraud and risk decisions in dollars instead of judging them on model metrics alone.
Fraud, risk and merchant analytics at Geidea
- Built real-time ML anomaly detection in Talend ETL on 1M+ daily payment transactions to catch fraud and settlement errors. It evolved into production fraud models.
- Segmented 40K+ merchants by MCC, interchange and chargeback behavior.
- Built merchant churn prediction and revenue forecasting (ARIMA, logistic regression) that informed a tiered retention strategy, cutting merchant churn from 35% to about 10% on a $5M+ revenue portfolio.
- Built settlement and onboarding data models covering 400K+ POS terminals.
- Automated collections pipelines that contributed to $8M+ in additional annual profit.
Public work on fraud and risk decisions
Each project runs end to end on generated data with a deterministic seed, with code and a full write-up on GitHub.
- Merchant freeze economics: 60,000 merchants; the 98%-recall freeze policy nets -$9.1M versus freezing nobody; cutting review time from 3 days to 1 is worth +$4.5M, more than a 2-point AUC gain (+$3.3M).
- Queue economics for a fraud and compliance review team: Alert recall 85.1% to 80.0%, dollar recall 76.8% to 93.0%, total cost -54%.
- Ledger integrity under agent-speed traffic: Netting hid $32,732 (46%) of $71,570 gross error; a retry with a regenerated idempotency key was caught 0 of 53 times and is 93% of undetected dollar-hours.
- Marketplace payout reconciliation: 122,810 orders; daily checks fire 31 of 31 days on a clean book; a fee-on-tax bug passes all reconciliation until fees are recomputed from policy.
What carries over
The common thread is that a risk model only matters through the decision it drives. In merchant-freeze-economics, the policy that catches 98% of bad actors is the worst one on the board, and cutting review time from 3 days to 1 is worth more than a 2-point AUC gain. In fraud-queue-economics, letting alert recall fall while dollar recall rises cuts total cost by 54%.
Tools
Python, SQL, scikit-learn, XGBoost, CatBoost, ARIMA, Talend, Oracle, Tableau, dbt, Snowflake.
Contact
Email sabasharat.ali@gmail.com, or find me on LinkedIn and GitHub. Full history on the experience page.
Open to roles in New York or remote US.