Computer Science Graduate | Aspiring Machine Learning Engineer & Data Analyst
Focused on practical Python + SQL systems, empirical evaluation, auditable data pipelines, and avoiding silent production pitfalls like target leakage and temporal distribution shift.

An empirical machine learning case study examining why telemarketing subscription models fail when deployed, built on 41,188 records from the real UCI Bank Marketing dataset.

• Leakage Prevention: Explicitly excludes duration (unknown before a call is answered) and enforces strict input validation with Pydantic v2 (extra="forbid" rejects duration with HTTP 422).
• Time-Ordered Split: Evaluates performance across chronological blocks (Train: May–Nov 2008, Val: Nov 2008–Jul 2009, Test: Jul 2009–Nov 2010) instead of randomized cross-validation.
• Model Governance: Analyzes the real-world negative result where temporal distribution shift causes the baseline model to fail validation, documenting why it must not be deployed.
• API & Surveillance: Tested FastAPI endpoint returning exact linear log-odds decompositions ($z = w_0 + \sum w_i x_i$) alongside standard distribution drift alerts.

An auditable analytics repository analyzing transaction integrity, customer decile concentration, and repeat purchasing on 541,909 raw invoice records from the UCI Online Retail dataset.

• Auditable SQL Models: Uses SQLite Common Table Expressions (CTEs), NTILE deciles, and window functions to reconcile £9.75M net signed sales across 4,371 identified customers.
• Reversal & Concentration Audits: Uncovers an 8.40% cancellation ratio and reveals that the top spend decile generates 60.14% of identified net revenue.
• Interactive Dashboard: Streamlit application backed by automated AppTest regression coverage.
• Executive Memo: Includes a structured decision memo distinguishing descriptive findings from causal claims and outlining reconciliation steps for enterprise stakeholders.

• GitHub: @ariffahmi0021
• Repositories: bank-campaign-ml · retail-cohort-analytics