ÖFK
Menu

Professional experience

MLOps & Analytics Engineer

Built and operated production ML, data and Generative AI platform capabilities across analytics workloads.

Company
Fibabanka
Team
Analytics Center of Excellence
Period
Mar 2023 — Mar 2026
Location
Istanbul, Türkiye
View as Markdown
Experience3+ YEARS
Production AI/MLProfessional experience across ML, data and GenAI systems.
Lifecycle10+ ML MODELS
Deployed & operatedProduction ML lifecycle ownership across multiple analytics use cases.
Scale~9M CUSTOMERS
Customer-level tablesBatch and near-real-time pipelines consolidating many source tables into customer-level training, inference and analytics tables.
Performance75% RUNTIME REDUCTION
Core ETL workflowSequential Oracle ETL re-architected with modular parallel dbt: 120 → 30 min.

Selected impact

Platform work through outcomes.

Stories are organized around engineering changes and measurable effects rather than a conventional responsibility list.

Impact / 01

Call-Center Intelligence

Call recordings held operational signal that nobody could query: reviewing them was manual, so the content never reached reporting.

Built an audio intelligence pipeline from transcription and diarization through structured operational outputs.

Audio pipeline throughput

~9,000 recordings / day

Production call-center pipeline

Daily recording volume processed end to end: transcription with WhisperX, timestamp alignment, speaker diarization, then LLM summarization into structured fields that feed operational KPIs.

  • WhisperX
  • Timestamp alignment
  • Speaker diarization
  • LLM summarization
  • Structured outputs
  • Operational KPIs

Impact / 02

ML Platform

Models were reaching production faster than there was a repeatable way to validate, promote and monitor them.

Owned deployment and lifecycle management for 10+ production ML models.

Production model ownership

10+ production ML models

Full lifecycle ownership

Ownership covered the lifecycle rather than a single handoff: validation, versioning, promotion, serving, retraining and monitoring across multiple analytics use cases.

  • Validation
  • Versioning
  • Promotion
  • Serving
  • Retraining
  • Monitoring

Impact / 03

AI Infrastructure

Generative AI prototypes validated on Azure could not stay there — customer data required the models to run inside the bank.

Moved validated Azure prototypes to on-premises GPU infrastructure for private open-source model serving.

  • Azure
  • On-prem GPU
  • llama.cpp
  • Ollama
  • LangChain

Impact / 04

Data Platform

The core ETL ran as sequential Oracle procedures, so the workflow's runtime grew with every addition and left no room in the batch window.

Re-architected sequential Oracle ETL workflows into modular parallel dbt pipelines.

Core ETL runtime

120 → 30 min

Sequential Oracle → modular parallel dbt

End-to-end runtime of the core workflow after splitting monolithic sequential procedures into modular dbt models that execute in parallel — a 75% reduction.

  • Oracle
  • dbt
  • ETL / ELT
  • Parallel execution

Impact / 05

Feature Platform

Every model rebuilt the same customer features in its own pipeline, so definitions drifted apart between use cases.

Built a centralized feature store and dimensional models used by 10+ production ML models.

Shared feature reuse

One definition per feature

Centralized store across 10+ models

Features and dimensional models were defined once and consumed by the production models, replacing per-model reimplementation of the same customer attributes.

  • Feature store
  • Dimensional modeling
  • Reusable features

Impact / 06

Data Scale

Training, inference and analytics all needed the same customer-level view, assembled from many separate source tables.

Engineered batch and near-real-time training, inference and analytics pipelines.

Customer-level coverage

~9M customers

Customer-level output tables

The pipelines consolidate many upstream source tables into customer-level tables covering the full ~9M customer base. This is the coverage of the produced tables, not a per-day record throughput.

  • Batch pipelines
  • Near-real-time
  • Training
  • Inference
  • Analytics

Impact / 07

Quality & Monitoring

Pipeline and model failures surfaced downstream as wrong numbers in reports rather than as an alert at the point of failure.

Implemented data quality, model validation and delivery workflows around production analytics systems.

Failure detection point

Checks in the pipeline, not the report

Data quality + model validation gates

Great Expectations and Deepchecks checks run as part of the Airflow pipelines and GitHub Actions delivery workflow, so data and model regressions are caught where they occur instead of being discovered in downstream analytics.

  • Great Expectations
  • Deepchecks
  • Airflow
  • GitHub Actions

Clear separation

Professional work shaped the engineering problems explored in public projects.

The portfolio projects are independent public engineering work. They are not Fibabanka source code or representations of internal systems.