Alihuseyn Ismayilov

PD Model for Retail Customers (Yelo Bank)

Built a Probability of Default (PD) model for retail customers within a one-month deadline mandated by the Central Bank of Azerbaijan.

Data Architecture & Preparation: Defined project scope, Critical Data Elements (CDEs), and feature list. Collected and validated data from Oracle SQL databases, verified against APIs, and cross-checked with business teams for accuracy.

Collaborated with Data Architects and Engineers to design and populate four datamarts:

• Target Datamart — Defines target variable and label structure.
• DPD Datamart — Includes Days-Past-Due (DPD) data from both Central Bank and internal systems.
• Contract-Based Datamart — Aggregates contract-level metrics (sum, min, max, avg) across 252 feature variations (4 aggregation types × 7 time windows × 3 portfolio groups × 3 data sources).
• Application Datamart — Contains operational data collected during customer loan applications.

Modeling & Evaluation: Applied Information Value (IV) and Weight of Evidence (WOE) for feature selection, followed by Logistic Regression with hyperparameter tuning. Evaluated multiple target definitions (15, 30, 60, 90 DPD) and loan age constraints (≥6 months, ≥12 months).

Performance Metrics:

• AUC: 0.78 (train), 0.74 (test), 0.71 (OOT)
• Precision/Recall: stable across datasets
• Approval Rate: optimized via business validation

Deployment & Validation: Validated by Internal Validation Team and ZYPLE, then deployed to Kubernetes for real-time scoring and monitoring.

Delivered the full pipeline — from data sourcing to deployment — in one month, setting a new internal benchmark for compliance-driven model delivery.

Cross-Sell Propensity Models (PASHA Bank)

Developed ML and rule-based recommendation models for FX, Foreign Payments, Loans, Salary Cards, and Factoring.

• Loans: used Positive-Unlabeled (PU) Learning to handle scarce negatives — suited for marketing where “not offered” ≠ “negative”.
• FX & Foreign Payments: RFM segmentation (Recency, Frequency, Monetary) with 5 labels (Active, Passive, About-to-be-Passive, Lost, Potentially Active).
• Other products: rule-based heuristics aligned with business criteria.

Data sources: Central Bank, transactional DWH, CRM. Built time-windowed features (3/6/12 months), defined CDEs, curated feature list, and materialized to a DWH datamart. Automated monthly in Dataiku (ETL → preprocess → model → DB write-back). Results power the internal RM Tool for sales prioritization.

Revenue uplift tracked via Tableau dashboards and PostgreSQL RM logs exceeded ₼1.2M AZN across 2023–2024.

Fraud Detection & Prevention (PASHA Bank)

Implemented a near real-time fraud detection framework combining unsupervised anomaly detection (Isolation Forest, Autoencoder) and supervised classification (Gradient Boosting, Logistic Regression) in Dataiku with PySpark for large data volumes.

Hybrid model performance (test env):

• Precision: ~88%
• Recall: ~82%
• AUC: ~0.91

The supervised model performed best and was promoted to production for continuous observation, reducing false positives by ~35% and analyst review time by ~25%.

News & NLP Analytics (PASHA Bank)

Built an end-to-end news intelligence pipeline for CRM and Counterparty Check. Scrapes trusted Azerbaijani sources every 10 minutes (static via requests + BeautifulSoup, dynamic via Selenium), assigns unique IDs and metadata, and stores to PostgreSQL.

Entity matching: For CRM, allowed 85–90% similarity using FuzzyWuzzy (Levenshtein). For Counterparty Check, applied strict normalization: dictionaries for legal forms like MMC (Mehdud Məsuliyyətli Cəmiyyət — Limited Liability Company) / LLC; transliteration rules (Paşa→Pasha); punctuation/spacing variants (m.m.c / m. m. c); and token co-occurrence (2–4-grams) using collections to capture names like “Bravo”, “Bravo MMC”, or “Bravo Mehdud Məsuliyyətli Cəmiyyəti”. Excluded lowercase common-word matches (e.g., “bravo”/congratulations).

Exposed an API through Dataiku for Counterparty Check and connected CRM dashboards so RMs see relevant news instantly. Result: RM search time reduced by 30–40 minutes per client; Counterparty Check usage-driven income up by ~20%.

Operator Email Efficiency Dashboard (PASHA Bank)

Measured First Response Time (FRT), Handling Time (HT), and operator efficiency, plus required staffing. Refactored legacy SQL with modular CTEs, efficient joins, and centralized date logic → ~90% faster queries.

Defined efficiency score: Efficiency = HandlingTime × log(1 + ConversationCount) to reward throughput while controlling extremes. Metrics honor business time (09:00–18:00), weekends, and holidays. Found Efficiency Factor ≈ 2.9, indicating potential 2–3× staffing optimization without SLA loss.

Dataiku Certificates

Choose a credential to open in a new tab:

Oracle SQL for Data Analysis — Certificate

Dean’s Honor — Gallery

This opens the Khazar honors lightbox. Click the university card in Education to browse all images.