Alpay Pasali
Alpay Pasali

Alpay Pasali

Backend • Senior • 30 years • Bursa, Turkey • UTC+03:00 • Work permits: Turkey • English: B2
Data Scientist
6+ years exp
20+ projects
Python
SQL
C#
Data Layer & Database: 5/10
API Design: 4/10
Scalability & Performance: 4/10
Invite to interview
Download CVCV
Overview
Technical skills
Projects
Timeline
Roles

Overview

Data Scientist / AI Engineer with 3+ years of software engineering experience and focused hands-on work in data science, machine learning, and AI-driven analytics. Skilled in Python, SQL, Power BI, ERP data workflows, REST API integrations, internal automation systems, and LLM-based applications. Experienced in transforming operational, production, warehouse, maintenance, and customer data into dashboards, analytical models, automation workflows, and AI-assisted business insights. Strong background in data preprocessing, feature engineering, relational data modeling, query optimization, API integration, and internal data application development.

Technical skills

Python
SQL
C#
Databases
PostgreSQL
MS SQL
Chroma
AI/ML
NumPy
Scikit-learn
ChatGPT
TF-Keras
Stable Diffusion
DALL-E
Jupyter Notebook
LLM Apps
DevOps
Rest API
Analytics
Power BI

Projects

ReAct Agent Chat Application
May 2026 to Jun 2026 1 Month

• Built a Streamlit-based AI chat application using LangChain’s ReAct agent architecture for tool-augmented reasoning and conversational AI.

• Integrated multiple LLM providers, including OpenAI, Google Gemini, and Groq, with selectable model options from a single user interface.

• Developed custom tools for web search, web scraping, and AI image generation using DuckDuckGo/Tavily, BeautifulSoup, DALL-E, and Stability AI.

• Implemented session-based chat history, real-time agent reasoning output, and configurable AI workflow options.

LLM-Based Textile Production Decline Detection
Dec 2025 to Mar 2026 3 Months

• Analyzed textile production data to detect product-level performance declines, anomalies, and operational trends over time.

• Performed exploratory data analysis, statistical analysis, and feature engineering to identify meaningful changes in production behavior.

• Used Large Language Models and prompt engineering to convert analytical outputs into natural-language business insights.

• Developed an AI-assisted analytics workflow that helped non-technical stakeholders interpret production KPIs,trends, and potential performance issues.

Factory Operations & Breakdown Analytics Dashboard (Power BI)
Jan 2026 to Jan 2026 0 Months

• Built a 10-page Power BI dashboard to analyze breakdowns, downtime, maintenance patterns, equipment failures, and department-level operational KPIs across multiple textile production units.

• Designed analytical data models and DAX measures for downtime analysis, breakdown frequency, equipment failure rates, shift-based performance, and maintenance prioritization.

• Transformed raw production and maintenance records into clear visual insights for operations leadership and non-technical stakeholders.

• Enabled factory teams to identify recurring equipment issues, compare department-level performance, and make data-driven maintenance decisions

Customer Segmentation & Churn-Risk Analysis
Nov 2025 to Jan 2026 2 Months

• Extracted and prepared customer data from internal business systems for segmentation and churn-risk analysis.

• Applied clustering algorithms to group customers based on behavioral and business-related patterns.

• Used machine learning and analytical insights to identify customer groups with higher churn risk and support data-driven business decision-making.

Timeline

Software Engineer • Middle
Surteks • Full-Time
Jul 2024 to Mar 2026 1 Year 8 Months Ivanovo In office

• Integrated ERP, API, production, warehouse, maintenance, and customer data into SQL-based reporting workflows, reducing manual data dependency and improving access to operational analytics.

• Developed AI-assisted analytics workflows for textile production data, using LLMs and prompt engineering to convert KPI trends, anomalies, and performance decline signals into business-readable insights.

• Optimized SQL queries, data models, and reporting processes, improving data retrieval performance by 40%.

• Developed internal data applications, REST API integrations, ERP automation workflows, and server-side processes to support real-time reporting and faster operational decision-making.

• Built a barcode-based warehouse tracking system to improve inventory traceability, product movement monitoring, and material accountability across warehouse operations.

• Applied machine learning clustering techniques to customer data for marketing-focused segmentation, customer profiling, and churn-risk analysis, enabling more targeted marketing and retention-oriented decision-making.

.NET Developer • Middle
Sifir Gibi Makine • Full-Time
Sep 2023 to May 2024 8 Months İzmit Fully remote

• Improved an ASP.NET-based application by refactoring existing code, resolving technical gaps, and increasing performance, maintainability, and reliability.

• Strengthened application security by improving access control, validation logic, and secure coding practices across key modules.

• Developed and supported REST API-related components to enable reliable data exchange, system integration, and workflow continuity.

• Optimized application workflows, database interactions, and query performance, improving system responsiveness by 60% while supporting a more modular and scalable architecture.

КГУ
Bachelor's Degree • Mathematical Support and Administration of Information Systems Comparable to Applied Mathematics and Computer Science
2019–2023 Kursk, Kursk Oblast
IT Specialist • Middle
Ekoteks • Full-Time
Jan 2023 to Jun 2023 5 Months Kursk In office

• Managed Windows Server environments and DNS configurations supporting data accessibility across the enterprise

• Developed systematic troubleshooting methodologies ensuring infrastructure reliability for data-dependent workflows

Yalova Üniversitesi
Associate's Degree • Paramedicine / Emergency Medical Services
2014–2016 Yalova, Turkey
Senior Backend Developer Confidence: Medium API Engineer
A backend-focused developer with a senior-level grasp of .NET platform patterns and reusable server-side libraries. The strongest proven skill is building cross-cutting backend infrastructure (repository abstraction, caching pipeline, exception middleware) as shown by EfRepositoryBase.cs and CachingBehavior.cs. There is limited public evidence of production-grade microservice contracts, migration histories, automated tests, or robust operational runbooks/metrics.
API Design
4/10
How well APIs are designed
API design shows deliberate error contract and pagination primitives (ProblemDetails, Paginate/PageRequest) and pipeline behaviors, but lacks explicit API versioning, idempotency key handling, and standardized rate-limiting.
Evidence
Core.Application/Request/PageRequest.cs
Core.Persistence/Paging/Paginate.cs
Core.CrossCuttingConcerns/Exceptions/HttpProblemDetails/BusinessProblemDetails.cs
Data Layer & Database
5/10
Working with databases
Data layer demonstrates thoughtful repository abstraction, EF-specific optimizations (AsNoTracking, dynamic includes), soft-delete traversal and dynamic filtering, showing schema/ORM awareness; however there is no visible migration history or tuned SQL/index definitions.
Evidence
Core.Persistence/Repositories/EfRepositoryBase.cs
Core.Persistence/Dynamic/IQueryableDynamicFilterExtensions.cs
Core.Persistence/Repositories/IAsyncRepository.cs
Scalability & Performance
4/10
Handling load and speed
Performance and scalability considerations are present via distributed caching with group invalidation and careful query options, but there is no evidence of queue-based decoupling, load-testing artifacts, or advanced connection/throughput tuning.
Evidence
Core.Application/Pipelines/Caching/CachingBehavior.cs
Core.Application/Pipelines/Caching/CacheSettings.cs
Core.Persistence/Repositories/EfRepositoryBase.cs
System Architecture
4/10
Overall system structure
The codebase uses clear module boundaries and cross-cutting pipeline behaviors (caching, validation, transactions, logging, exception middleware), indicating deliberate architecture for server-side components; microservice decomposition and inter-service contracts are not evidenced.
Evidence
Core.Application/Pipelines/Caching/CachingBehavior.cs
Core.CrossCuttingConcerns/Exceptions/ExceptionMiddleware.cs
Core.Application/Pipelines/Transaction/TransactionScopeBehavior.cs
Security & Auth
3/10
Protecting data and access
There is basic security awareness - centralized exception handling and validation pipeline are present - but no evidence of authn/authz patterns, token lifecycle, secrets rotation, input sanitization in web-facing tools, or dependency vulnerability audits.
Evidence
Core.CrossCuttingConcerns/Exceptions/ExceptionMiddleware.cs
Core.Application/Pipelines/Validation/RequestValidationBehavior.cs
Core.CrossCuttingConcerns/Exceptions/HttpProblemDetails/InternalServerProblemDetails.cs
Reliability & Observability
4/10
Stability and monitoring
Observability and reliability are addressed with structured logging integration and exception middleware; async methods accept CancellationToken, but there is limited evidence of retries/backoff, circuit breakers, or operational runbooks/metrics instrumentation.
Evidence
Core.CrossCuttingConcerns/SerliLog/LoggerServiceBase.cs
Core.Application/Pipelines/Logging/LoggingBehavior.cs
Core.CrossCuttingConcerns/Exceptions/ExceptionMiddleware.cs
Expertise
.NET• Senior
Backend AI & LLM• Middle
Industries
Artificial Intelligence• Middle
Data & Analytics• Middle
Recommendations
  • Lead development of backend platforms and shared libraries - middleware, repository patterns, caching strategies and consistent error contracts for teams using .NET.
  • Implement and harden service-level features such as API versioning, idempotency handling, and documented migration chains for database schema evolution.
  • Prototype LLM-backed features or tools (Streamlit/langchain integrations) while pairing with a security review to address input sanitization, timeouts and file-handling.
  • Add operational observability: structured metrics, traces, retry/backoff patterns and simple load tests to validate caching and query performance.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
alpaypasali ReAct_Agent_Chat_Application .Net-Core-Packages
Middle AI/ML Engineer Confidence: Medium ML Engineer
A developer at a middle level who builds end-to-end ML prototypes and multi-model LLM demos. The strongest proven skill is building applied ML workflows and prototypes, demonstrated by the complete Keras CNN training pipeline (data split, preprocessing, tf.data, model.fit with EarlyStopping/ModelCheckpoint) in the handwritten-letter notebook. What is not evidenced is production-grade MLOps, experiment tracking, custom model architectures or research-level contributions.
Model Architecture & Training
4/10
How well models are designed and trained
Correct, conventional model building and training with TF-Keras (Sequential CNN, Adam optimizer, categorical crossentropy, AUC metric). No custom architectures, custom loss functions or advanced training schedules.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: model = Sequential([...]) alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=metrics)
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Complete data preparation pipeline for image classification - stratified train/val/test split, normalization, one-hot encoding and tf.data batching. Standard but well-structured preprocessing and dataset creation.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: get_images_with_labels(...) alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: load_and_preprocess(...) and prepare_datasets(...)
Experimentation & Evaluation
3/10
How results are measured and tested
Basic experimentation and evaluation: training logs, EarlyStopping, ModelCheckpoint, plotting training history and final evaluation metrics. No experiment tracking (W&B/MLflow), systematic ablation, or multiple-run reproducibility artifacts.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: EarlyStopping and ModelCheckpoint used in model.fit(...) alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: plot_training_history(history...) and model.evaluate(...)
MLOps & Deployment
2/10
How models are shipped to production
Lightweight serving/prototyping evidence - a Streamlit-based ReAct agent front-end and saving Keras model as HDF5. Lacks deployment automation, model versioning, monitoring, CI/CD or container/orchestration artifacts.
Evidence
alpaypasali/ReAct_Agent_Chat_Application/react_chat.py: Streamlit app wiring and AgentExecutor invocation alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: ModelCheckpoint('best_model.h5', ...)
Computational Efficiency
2/10
How efficiently computing resources are used
Minimal efficiency work: use of tf.data and batching is present but there is no evidence of quantization, mixed precision, profiling, optimized kernels or distributed training.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: prepare_datasets(...) uses tf.data.Dataset alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: notebook logs show CUDA initialization message but no optimization code
Research Depth & Innovation
1/10
Depth of research and new ideas
No research-level innovation: the notebook implements a standard Conv2D-based CNN and common training techniques, with no novel layers, algorithms or reproduced paper-level contributions.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: model architecture is Conv2D/MaxPooling/Flatten/Dense blocks
Expertise
AI Agents & Agentic Workflows• Middle
Conversational AI & Chatbots• Middle
Technologies
DALL-E
Recommendations
  • Develop prototypical computer vision models and Kaggle-style image-classification pipelines (data preprocessing, training, validation, basic deployment).
  • Build multi-model LLM proof-of-concept apps and agent-driven demos (Streamlit front-ends, tool integrations, streaming callbacks) while hardening error handling and credential management.
  • Focus on adding reproducible experiment tracking and basic MLOps (W&B/MLflow, Docker + simple CI, model versioning) before taking on production ML infrastructure work.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
alpaypasali ReAct_Agent_Chat_Application Russian-Handwritten-Letter-Recognition-with-CNN
Middle Data Scientist Confidence: Medium ML Practitioner
A practical machine-learning practitioner at a solid middle level who delivers end-to-end experiments and simple LLM integrations. The strongest proven skill is building and training applied deep-learning models, evidenced by a full CNN training notebook with stratified splits, tf.data pipeline, early stopping and checkpointing. What is not evidenced is production-grade MLOps - unit tests, CI/CD, dependency pinning, experiment tracking or robust hyperparameter search are missing.
Statistical Rigor
3/10
Correct use of statistics
Uses appropriate evaluation metrics (accuracy, AUC) and stratified splits but lacks formal statistical tests, uncertainty quantification, or multiple-comparison controls.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: train_test_split(..., stratify=labels) and AUC metric in model.compile alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: evaluation prints Loss/Accuracy/AUC but no confidence intervals or hypothesis tests
Data Wrangling & Cleaning
5/10
Preparing and cleaning data
Clear data ingestion and preprocessing utilities, normalization, one-hot encoding and tf.data pipelines with stratified dataset splitting and explicit file-to-label mapping.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: get_images_with_labels() builds file->label pairs from CSV alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: load_and_preprocess() normalizes images and prepare_datasets() creates tf.data.Dataset
Exploratory Analysis & Visualization
5/10
Exploring and visualizing data
Provides exploratory visuals and written narrative - image display, random samples, training/validation curves and prediction visualizations with color-coded correctness, but analysis is focused on model performance rather than deeper data-driven storytelling.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: display_images() and show_random_examples() for EDA alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: plot_training_history() and show_predictions() visualizations
Predictive Modeling
5/10
Building models that predict
Implements an end-to-end CNN training pipeline with sensible architecture, metrics, early stopping and checkpointing; lacks systematic hyperparameter search, robust CV (beyond a single stratified split), or advanced error analysis.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: model = Sequential([...]) with Conv2D/MaxPooling/Dense and model.fit(...) alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: EarlyStopping and ModelCheckpoint callbacks used
Business Insight & Impact
Turning analysis into business value
Not evidenced in public code
Reproducibility & Notebook Hygiene
3/10
Clean, repeatable analysis
Some reproducibility hygiene is present (fixed random_state in splits, model checkpointing) but environment pinning, dependency manifests, experiment tracking, and data/version control are missing.
Evidence
alpaypasali/Russian-Handwritten-Letter-Recognition-with-CNN/russian-handwritten-letter-recognition-with-cnn.ipynb: train_test_split(..., random_state=42) alpaypasali/ReAct_Agent_Chat_Application/customtools.py and react_chat.py: use of load_dotenv() and os.getenv for API keys (env-based config) and presence of notebook meta ('# NOTEBOOK META')
Technologies
LLM Apps
ChatGPT
Stable Diffusion
Jupyter Notebook
Scikit-learn
NumPy
TF-Keras
Recommendations
  • Harden ML projects for production: add dependency pinning (requirements or environment.yml), CI tests, and a reproducible runbook or Dockerfile for training and inference.
  • Introduce experiment tracking and model versioning (MLflow, Weights & Biases, or DVC) and automated evaluation to enable repeatable comparisons and rollback.
  • Add systematic hyperparameter tuning and robust cross-validation or time-aware splitting where applicable, plus more error-analysis (confusion matrices per class, per-class metrics) to diagnose failure modes.
  • For LLM apps, add input validation, tool sandboxing, and explicit output-sanitization to reduce injection risks and to separate model orchestration from user-facing logic.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
alpaypasali ReAct_Agent_Chat_Application Russian-Handwritten-Letter-Recognition-with-CNN