Experience & advisory
Work
Since September 2020, my independent consulting work has grown from a desktop forecasting workbench into model-serving systems, team leadership, and mentoring, alongside my engineering roles. I’m a Senior ML Engineer and hands-on Technical Lead leading AI R&D at Universal Software; Moka is my only active consulting client. Below are my professional roles and past and current engagements, with dates and tech stacks. For a concise overview, see my CV.
Professional roles
3 positions-
Technical Lead, AI R&D
Lead the hands-on AI R&D roadmap, architecture reviews, delivery planning, and stakeholder coordination; mentor up to 5 engineers building ML products for municipalities, ministries, and utilities.
Highlights & technologies
- Architected MdsAI, a human-in-the-loop support intelligence platform for 84 municipalities over 132K support events / 29K tickets. Evidence retrieval reached 97% Recall@5 offline; on-premises answer drafts from fine-tuned Gemma 4 / Qwen models served with vLLM, plus automated routing, helped cut median first-response time by 50%.
- Co-lead a TÜBİTAK 1832 national R&D grant with hydraulic engineers; designed attention-based seq2seq river-discharge forecasting with event-based backtests and FastAPI serving for 8 gauging stations (Teus).
- Led the technical work on Horizon Europe consortium proposals for flood and fire prediction, translating domain requirements into research plans and system designs.
- Python
- PyTorch
- LLMs
- RAG
- Embeddings
- seq2seq + Attention
- FastAPI
- Team Leadership
-
AI/ML Engineer → Senior AI/ML Engineer
Built and deployed forecasting, computer vision, and predictive-maintenance systems for municipal utility networks and industrial clients.
Highlights & technologies
- Shipped a real-time Enerjisa field-photo validation pipeline combining YOLOv8 detection with blur, duplication, and empty-frame checks, reducing invalid submissions reaching inventory review by 30%.
- Designed IoT-driven waste-collection routing that converted live bin-fullness and fire/flood sensor events into priority routes, reducing planned distance by 18% in historical simulations.
- Built predictive-maintenance classifiers for Enerjisa assets and CNC assembly lines, identifying 82% of failures within a 7-day warning window.
- Developed LSTM pipelines for municipal water-demand and industrial gas/electricity forecasting, improving MAPE by 21% relative to seasonal-naive baselines; also built TF-IDF/embedding-based citizen-complaint routing for client municipalities.
- Python
- YOLOv8
- LSTM
- scikit-learn
- TensorFlow
- Keras
- Heuristics
- IoT Telemetry
-
AI/ML Engineer (Part-time Contract)
Built production recommender systems for commercial consumer platforms, combining multimodal and collaborative-filtering approaches.
Highlights & technologies
- Built a multimodal recommender for the online art gallery 'Collectors', matching artwork to users through textual profiles and CNN-extracted visual preferences.
- Developed a collaborative-filtering engine for the audiobook app 'Dinlebi' using matrix factorization over user interaction history, with an onboarding flow designed to mitigate cold start.
- Increased recommendation CTR by 14%; both the Collectors and Dinlebi systems remained in production for 3+ years.
- Python
- CNNs
- Transformers
- Matrix Factorization
- Collaborative Filtering
- Content-Based Filtering
- Deep Learning
Flagship engagements
7 clients-
Inveon
Built semantic product search and a mixture-of-experts classifier reaching 91% top-1 accuracy across 50+ subcategories, in production since 2023. Also delivered multimodal similarity/recommendation systems, order cancellation/return prediction APIs, and shipment-to-delivery analytics with carrier-specific forecasts and comparisons against actual delivery times.
- Python
- FastAPI
- Streamlit
- PyTorch
- Scikit-learn
- Image Classification
- Deep Learning
- Semantic Search
- Recommender Systems
- Cancellation/Return Classification
- Delivery Analytics
-
Moka
Build churn/lifetime-value scoring, price-quote prediction, and financial forecasting behind FastAPI with validated feature contracts, on Azure, Docker, and Kubernetes. The merchant-churn model reached 0.93 F1 / 0.95 ROC-AUC in nested cross-validation. Serving work includes batch inference with missing-feature checks and wallet, distributor, and POS forecasting endpoints.
- Python
- FastAPI
- Streamlit
- Docker
- Kubernetes
- Azure
- Git
- Fintech Analytics
- Time-Series Forecasting
- Churn/LTV
-
Ikas
Built call-center planning APIs that turn call-volume and handling-time forecasts into Erlang-based staffing requirements, alongside demand and revenue forecasting services. Delivered Tkinter MLStudio apps, Streamlit tools, and FastAPI inference endpoints returning agent requirements, service-level estimates, and queue/utilisation measures.
- Python
- Tkinter
- Streamlit
- Polars
- FastAPI
- Time-Series Forecasting
- Revenue Prediction
- Demand Forecasting
- Call-Center Staffing / Erlang
-
Datakod
Designed a broad business intelligence and analytics system for e-commerce distributor channels. Features include market basket association rules (FP-growth, Apriori), RFM metrics, anomaly detection, sales forecasting, and customer clustering.
- Python
- Pandas
- Scikit-learn
- Keras
- Statsmodels
- Association Rules (Apriori/FP-growth)
- RFM Analytics
- Clustering
- Anomaly Detection
- Sales Forecasting
-
Badem
Built concrete-strength regression applications, doctor/customer churn models, lifetime-value estimates, and segmentation reports. Delivered Tkinter MLStudio applications and Streamlit tools.
- Python
- Tkinter
- Streamlit
- Pandas
- Scikit-learn
- Regression
- Classification
- Churn/LTV
- Segmentation
-
Albert Solino
Studied BERT-based HR candidate identification, built a database-connected churn/LTV tool and customer/email segmentation pipelines, and mentored two mid-level engineers who built a server-rendered Flask MLStudio application.
- Python
- Sentence-Transformers
- PyTorch
- Streamlit
- Flask
- XGBoost
- NLP / HR Analytics
- CRM Integration
- Mentoring
-
Smartiks
Led a three-person AI/R&D team on predictive modeling, segmentation, and drone-based bird detection, and wrote the technical parts of ITEA consortium proposals. Modeling work included BMR prediction for health clients and campaign-oriented customer clustering with K-Means, DBSCAN, and CLARA.
- Python
- TensorFlow
- Keras
- Yolo
- OpenCV
- XGBoost
- Regression
- Clustering (K-Means/DBSCAN/CLARA)
- Computer Vision
- ITEA Project Leadership
Supporting engagements
15 clients-
Barkosoft / Hesapcini
Developed operational demand forecasting engines and API integrations for Hesapcini pre-accounting SaaS and WMS solutions.
- Python
- Pandas
- Scikit-learn
- Keras
- Time-Series Forecasting
- WMS & Pre-accounting APIs
-
KolayBi
Designed churn prediction, customer lifetime value models, and demand forecasting backend APIs for SME finance SaaS platforms.
- Python
- Pandas
- Scikit-learn
- Keras
- XGBoost
- Churn/LTV Modeling
- Forecasting APIs
-
Kriko
Modeled organic marketing and network traffic forecasting, presenting forecasts via a customized Streamlit end-user dashboard.
- Python
- Pandas
- Streamlit
- Keras
- Network Traffic Forecasting
- Marketing Analytics
-
Robokobi
Built financial forecasting workflows for CRM and SME business-management software with Tkinter and Streamlit tools. Implemented ONNX Runtime CPU inference with saved model configuration, scaling, differencing, and recursive forecast reconstruction.
- Python
- Tkinter
- Streamlit
- Keras
- ONNX Runtime
- Financial Forecasting
- SME Business Workflows
-
Vardabit
Created bid quote prediction models and automated sales forecasting pipelines, utilizing Tkinter model-builders and Streamlit web applications.
- Python
- Pandas
- Scikit-learn
- Tkinter
- Streamlit
- Keras
- Quote Prediction
- Sales Forecasting
-
Unsped
Built import/export logistics analytics, customer churn classification, and CLTV regression, alongside R&D proposal work under TÜBİTAK 1511.
- Python
- Pandas
- Scikit-learn
- XGBoost
- Tkinter
- Streamlit
- Logistics Analytics
- Churn/CLTV Modeling
- TÜBİTAK 1511 Proposal
-
Protel
Mentored and led a team of five data scientists and engineers on developing large-scale restaurant demand forecasting algorithms using time-series and gradient boosting.
- Python
- ARIMA
- LightGBM
- XGBoost
- Demand Forecasting
- Technical Leadership
-
Teknorot
Built automotive export demand forecasting engines for steering and suspension spare parts. Shipped a custom Tkinter MLStudio application managing 10-20 distinct model configurations.
- Python
- Tkinter
- Keras
- TensorFlow
- Automotive Spare-Parts Logistics
- Time-Series Forecasting
-
Comdata
Pioneered call-center call volume and handling time forecasting pipelines. Implemented an autoencoder-based anomaly detection workflow to scrub historical training data.
- Python
- Tkinter
- Keras
- TensorFlow
- Autoencoders
- Call-Center Analytics
- Anomaly Detection
-
SefaMerve
Developed a retail demand forecasting system for new products using CNN-based image similarity matching to historical product data, mapping forecasts based on style resemblance.
- Python
- Tkinter
- CNN
- Image Similarity
- Retail Demand Forecasting
-
Odeal
Conducted merchant transaction behavior clustering, customer churn analysis, and MLP modeling, publishing results via Streamlit prototypes.
- Python
- Scikit-learn
- MLP
- Streamlit
- Fintech Analytics
- Merchant Transaction Clustering
- Churn Analysis
-
Gtech
Mentored three mid-level data scientists in machine learning and software engineering; they built a desktop MLStudio tool for their company as the outcome of that work.
- Python
- MLStudio
- Forecasting Workbench
- Mentoring
-
Nano
Provided corporate training on machine learning and deep learning, guiding internal developers on building a custom C# MLStudio tool for regression and forecasting.
- C#
- MLStudio
- Regression & Forecasting
- Corporate Training
-
HepsiEmlak
Scoped and drafted R&D Center registration and TÜBİTAK grant applications. Designed algorithms for listing recommendations, rental valuation, and duplication checking using CNN image matching.
- CNN
- Recommendation Systems
- Image Duplicate Detection
- R&D Grant Writing
-
Hipposoft
Modeled customer churn and demand forecasting feasibility frameworks for various SME technology services.
- Python
- Scikit-learn
- Churn Prediction
- Feature Engineering
Additional engagements
8 clients-
Ayasof
Generated demand and sales forecasting pipeline reports for retail operations.
- Python
- Pandas
- Regression Forecasting
- Retail Operations
-
CukurovaMakine
Designed and delivered an intensive 8-module machine learning and deep learning curriculum for industrial engineering teams.
- Python
- Industrial ML/DL Curriculum
- Corporate Training
-
JForce
Conducted demand forecasting feasibility analysis for restaurant operations and delivered ML/DL training to more than ten engineers.
- Python
- Demand Forecasting Feasibility
- Engineer Training
-
SmartPulse
Researched energy/SCADA and weather-based load forecasting feasibility using Extreme Learning Machine (ELM) networks.
- Python
- Extreme Learning Machine
- Energy Forecasting
- SCADA Analytics
-
Storespy
Developed app store download and ranking prediction models using minimum Redundancy Maximum Relevance (mRMR) feature selection.
- Python
- mRMR
- XGBoost
- Deep Learning
- App Store Ranking Prediction
- Feature Selection
-
VTC
Created LSTM deep learning time-series models for energy forecasting and imbalance cost reporting.
- Python
- LSTM
- TensorFlow
- Keras
- Energy Trading
- Time-Series Forecasting
-
Ortus
Produced feasibility reports on revenue model analysis, incorporating customer clustering and mRMR feature selection.
- Python
- Customer Clustering
- mRMR Feature Selection
-
PayTR
Conducted fraud modeling evaluation, creating classification testbeds and metrics reports for risk analysis.
- Python
- Scikit-learn
- Fraud Detection
- Risk Classification
Reusable consulting tools
1 tools-
MLStudio
Built a reusable desktop forecasting workbench used across 30+ organizations, cutting time to a first validated model from 10 working days to 2. Later rewrote it in Streamlit/PyTorch with time-aware validation and versioned model bundles.
- Python
- Tkinter
- Streamlit
- PyTorch
- Polars
- scikit-learn
- Forecasting Workbench
- Model Validation