AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
Staff Data Scientist/ML Engineer at Enerjisa
A resilient leader, transforming needs of stakeholders into added-value by designing & developing data-driven solutions, through an effective process based on involvement, communication and transparency. Please visit my portfolio for the details of experience, projects and more. Portfolio: https://erelcan.github.io Leadership & Management ⚡ Effective team-building & team-branding for fast-paced environments ⚡ Casting vision to common goals and establishing a sense of purpose in the team ⚡ Stakeholder alignment through active involvement and action-yielding communication ⚡ Diplomatic management for multi-partnered national & international projects ⚡ Coaching & mentorship supported delegation Data Science & ML Engineering ⚡ Designing & developing robust ML/DL solutions tailored for domain and use-cases ⚡ Customizable, extensible and composable ML-automation for domain analysis, experimentation and reporting ⚡ Agentic-AI automatization through tool-using, multi-step, self-correcting agents with customizable mind and memory ⚡ Reliable decision-making solutions based on sound research, statistical evaluation and ExplainableAI ⚡ Framework-free paper-to-code development competence Data Engineering ⚡ Designing data architecture for use-case requirements and constraints ⚡ Developing & automating data ingestion and processing pipelines ⚡ Developing on-demand, batch and streaming solutions with broad data-variety (i.e. strucured, text, image, video, vectors etc.) ⚡ Optimizing batch processing/inference through distributed pipelines across CPUs/GPUs. Software Engineering ⚡ Building ML/DL solutions as re-usable, composable and extensible software ⚡ Optimizing (LLM-)serving and distributed processing; along with enhanced resource management, auto-scaling and fault-tolerance ⚡ Developing front-end & back-end for mobile, web and desktop ⚡ Port
Koç University
Master’s Degree, Computer Science and Engineering
January 1, 2011 – January 1, 2013
Koç University
Bachelor’s Degree, Computer Engineering
January 1, 2006 – January 1, 2011
Özel Denizli Fen Lisesi
High School, Science
January 1, 2002 – January 1, 2006
Enerjisa
Staff Data Scientist/ML Engineer
September 1, 2020 – Present
Ankara, Türkiye
HAVELSAN
Data Science Team Lead
January 1, 2019 – September 1, 2020
HAVELSAN
Software Engineer/Machine Learning Specialist
May 1, 2016 – January 1, 2019
TSK (Türk Silahlı Kuvvetleri)
Engineer Ensign
December 1, 2014 – October 1, 2015
Türkiye
Koç University
Research & Teaching Assistant
September 1, 2011 – August 1, 2013
Istanbul, Istanbul, Türkiye
HAVELSAN
Summer Intern
July 1, 2010 – July 1, 2010
Ankara, Türkiye
Valensas
Summer Intern
July 1, 2009 – July 1, 2009
Istanbul, Istanbul, Türkiye
Automated Video Understanding
January 1, 2026 – January 1, 2026
Video analysis takes huge amounts of time in operations. AI-powered video analysis facilitates the identification and prioritization of misapplications, enabling more timely response actions. Data Engineering: Implemented vector-db abstractions over FAISS and QDRANT, enabling (bulk) insert, update, delete and search. Developed QDRANT-based KNN by utilizing payload mechanism. ML Engineering: Designed and developed multiple strategies (combinations of embedding-approach x training-approach) for video classification along with prompt-based visual descriptors. Via configuration-controlled prompt and template generation, enabled time-stamp/interval based video descriptions, key information extraction and question-answering over videos. Automated embedding batch pipelines/jobs to evaluate the effect of various representations (video->embedding, video->text->embedding, video+text->embedding etc.). Developed ANN-based KNN for baseline along with transfer learning with stronger models in order to enable video classification.
Automated Natural Language Understanding (NLU)
January 1, 2025 – January 1, 2025
In electricity distribution grid operations, lots of text-based data is collected on the field and operations. To extract/compare information over such data manually consumes great amount of time and effort. Moreover, time-constraints brings risk of mis-evaluation. By developing automated NLU, hundreds of thousands of hours are saved for operations. Data Science: Designed and developed outlier detection mechanisms to clean previously annotated data. Applied active learning by implementing various sampling strategies (uncertainty-sampling, diversity-sampling etc.) to reduce annotation effort for domain experts. Conducted carefully designed systematic experiments to evaluate solution performance. ML Engineering: Implemented various text classification solutions based on few-shot learning, transfer-learning and fine-tuning. Built focus-adjusting cluster analysis tool based on text-embeddings (via HG-transformers and VL-models). Developed batch processing Ray/local jobs for HG-transformers and VL-models (both local and online).
MCP-Toolbox
January 1, 2025 – January 1, 2026
As operation teams have large variety of tasks, single agent may not handle all by itself. Hence, delegating work to tools and co-agents may enable better coverage of tasks and enhanced performance. Designed and developed MCP-Toolbox to integrate any required tool to AI-Assistant as plug-ins. Tool-Set/Co-Agents - Database Analyser (Co-agent): Gets request from supervisor agent, generates and executes sql query. - Database Analyser: Given (generated) sql query, executes it on db (there are various tools per db). - Excel Analyser: Given excel id(s) (managed by mind) and generated sql query, inserts csv to in-memory db and executes the query. - CSV Analyser: Given csv id(s) (managed by mind) and generated sql query, inserts csv to in-memory db and executes the query. - Image Descriptor (Co-agent): Given image id(s) (managed by mind), describes what is in the image. - Image Analyser (Co-agent): Given video id (managed by mind) and required (generated/static) prompt for analysis, extracts information from image. - Video Analyser (Co-agent): Given video id (managed by mind) and required (generated/static) prompt for analysis, extracts information from video. - Video Editor: Given video ids (managed by mind); cuts/merges video for several intervals and/or merges multiple videos into single video. - Web Searcher: Given request, searches web and returns relevant information. - Chart Creator: Given data and plot-meta, draws a plot, saves it to object store and returns resulting object id. There are also tools/co-agents tailored for specific requirements of operations.
Auto SQL Generator/Executor
January 1, 2025 – January 1, 2026
Operations teams handle many requests that become time-consuming when constrained by legacy ERP systems, especially without integrated SQL execution or sufficient SQL expertise. Integrated sql generation and execution capabilities to AI-Assistant to reduce operational costs. Data Engineering: Developed SQLite and SAP HANA integrations along with required data processing. Built auto-query execution as mcp tools with integrated security checks. Data Science: Led business teams to create benchmark data for Agentic SQL Generation. Executed experiments and evaluation on single agent and co-agent approaches on benchmark data. ML Engineering: Applied prompt-engineering for enhanced query generation performance and reducing unnecessary query execution. Developed N-step tool usage for requests requiring multiple sql execution and chart plotting.
Automated LLM-based Information/Feature Extraction
January 1, 2025 – January 1, 2026
On-premise LLM-serving infrastructure paved the way for many applications in regulatory domains of the company. By abstracting and representing recurring parts of applications (i.e. prompt-generation, LLM-management, templating, processing, deployment etc.) via DSL; time-to-market is optimized for fast MVPs, baselines and even production-level deployments. ML Engineering: Designed and developed automated (DSL -> End-to-end) LLM-based KIE library. Designed & developed fault-tolerant, distributed, pace-adjustable batch-processing jobs/pipelines.
The AI-Assistant
January 1, 2025 – January 1, 2026
Operation teams require collecting and evaluating information from many sources to reply requests or write reports, for daily tasks. AI-Assistant brings operational efficiency and enhanced quality for such operational tasks, as number of recurring tasks are high and time-constraints are tight. Data Engineering: - Developed file server for serving object store (s3) data via required protocols (i.e. http etc.), to be utilized in assistant markdown generation when injecting produced files (images, videos etc.) into assistant's response. - Designed and developed file upload/attachment mechanism for both user prompt and assistant-response. Implemented session-based id generation, s3 object upload/download and content conversion mechanisms. ML Engineering: - Designed and developed customized on-premise, multi-step tool using, multi-media enabled AI agents. - Designed the concept of mind which customizes prompts not only with raw message history but also with attachments, and derived information from them. - Mind concept brings rich and on-demand sub-contexts by managing references to previous files and derived content along with dynamic prompt injections. - Designed and developed memory concept to manage messages and attachments dynamically for mind and agentic-steps. - Designed and developed generic agents having varying behaviors (one-step agent, self-chained-agent, n-step agent etc.). - Integrated custom MCP-Toolbox to agents and enabled multi-step tool-usage with error correction. Software Engineering: - Provided abstractions for core components (i.e. llm-clients, agents, minds, memories etc.) to enable plug-in based incremental development. - Designed and developed demo front-end (i.e. chat, file upload, multi-media attachment viewers, memory/mind observers etc.) for fast MVP. - Enabled streaming token handling for clients and integrated to UI components. - Improved maintainability and flexibility by decoupling configuration/prompts from agent-core.
ML-Ops
January 1, 2025 – January 1, 2026
Due to regulations in energy domain, range of applications for cloud integration is limited. Therefore, on-premise development and productions environments are MUST for regulated domains. Revolutionized Enerjisa's infrastructure to pave the way for many applications to enhance operational efficiency and quality. Led the architecture design of data-science/AI clusters. Integrated, configured and tested Ray on Kubernetes(+Rancher) along with GPU setup. Optimized LLM-Serving with Ray + vLLM. Automated job submission mechanism, scheduling and monitoring with Ray (+Grafana/Prometheus). Led limited DevOps engineer resource for building and testing fault-tolerance, high availability, API-gateway, MIG-setup and containerized development. Critical tests: Ray caching (env, workspace etc.), dynamic request batching, resource utilization, GPU/Node-specific execution, Multi-LLM serving, customized service deployment, auto-scaling, distributed batch processing.
Document Structure Extraction
January 1, 2024 – January 1, 2024
Retrieval Augmented Generation (RAG) is a low cost option for low-resource languages and domains. Document structure extraction and enriched embedding strategies can improve RAG-based solutions. ML Engineering: Developed document (i.e. PDF) section/sub-section, table and plot extraction solutions for both text-based and scanned documents (with OCR, Table Transformer)
Web Portfolio (Self-Improvement)
January 1, 2024 – January 1, 2024
Developed a web portfolio on top of a excellent open source project. Enabled resume-to-filtered-project-routing, through tag-supported routing and tag-search mechanism. Also added manual tag search with suggestions. Customized CSS and SVG contents for better consistency. Customized section structure for better story-telling. Added markdown support for description cards. Generated (logo) images via diffusers (Generative AI).
Information Extraction on Inventory Label Plate (KIE)
January 1, 2024 – January 1, 2026
In an electricity distribution grid, maintaining critical inventory (e.g. transformers etc.) data is crucial for continuous electricity supply and for optimizing a variety of processes (e.g. maintenance, repairment, investment etc.). Due to several reasons (e.g. hand-over areas, outsource quality etc.), some useful properties of inventories may be missing or misleading. By extracting information from operation images, data-enhancement is achieved in short-time by saving huge costs in terms of operation-hours. Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, project management Data Engineering: Designed & developed image pipelines for a variety of data sources (e.g. NAS, REST Service, gRPC Service) ML Engineering: Designed both Text Detection & Recognition based and Visual Query Answering based solutions. Developed critical post processing (e.g. voting mechanism, out-of-scope information detection, match-confidence scoring, uniquification etc.) to ensure robustness. Developed object detection based QR localization solution for QR detection capabilities. Integrated auto-annotation support for domain-experts during annotation of benchmark datasets for evaluation.
Electricity Grid Connectivity Analysis
January 1, 2023 – January 1, 2023
Electricity distribution grid contains millions of connected inventories. While adding new inventories (or re-shaping sub-grid) due to investments, maintenance or hand-overs, drawings are ingested manually. As this is a cumbersome manual work, open to human and tool errors, there can be connectivity/flow errors in the data. To detect connectivity errors, conducted graph analysis and prepared the ontology of the graph. Led domain experts to identify mis-connections, along with providing outlier analysis support. Leadership & Management: Stakeholder management, coaching & mentorship, project management Data Engineering: Converted raw tabular data into graph representation. Data Science: Conducted domain analysis and prepared ontology of the connectivity graph. Applied outlier analysis on connectivity data to detect mis-connections in the graph.
Information Extraction on Inventory Label Plate (KIE)
January 1, 2023 – January 1, 2024
In an electricity distribution grid, maintaining critical inventory (e.g. transformers etc.) data is crucial for continuous electricity supply and for optimizing a variety of processes (e.g. maintenance, repairment, investment etc.). Due to several reasons (e.g. hand-over areas, outsource quality etc.), some useful properties of inventories may be missing or misleading. By extracting information from operation images, data-enhancement is achieved in short-time by saving huge costs in terms of operation-hours. Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, project management Data Engineering: Designed & developed image pipelines for a variety of data sources (e.g. NAS, REST Service, gRPC Service) ML Engineering: Designed both Text Detection & Recognition based and Visual Query Answering based solutions. Developed critical post processing (e.g. voting mechanism, out-of-scope information detection, match-confidence scoring, uniquification etc.) to ensure robustness. Developed object detection based QR localization solution for QR detection capabilities. Integrated auto-annotation support for domain-experts during annotation of benchmark datasets for evaluation.
Image Quality Analysis (IQA)
January 1, 2023 – January 1, 2024
Operation images are used as proof for regulations. Moreover, they also carry valuable information for a wide range of operations (i.e. inventory, maintenance, repairment, investment). Hence, ensuring image quality on the field is crucial. Therefore, developed IQA solutions and also improved operations' processes (through standardization, image enrichment with EXIF tags etc.). Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management Data Engineering: Developed pipelines for image ingestion and sampling over a set of metadata. (i.e. Oracle, REST Services). ML Engineering: Developed a variety of IQA solutions (i.e. statistical, classical ML, deep learning) for technical (i.e. blurred, extreme/low illumunation, occluded lens) and semantic image quality. Led domain-experts to annotate benchmark datasets for evaluation.
GenAI Frenzy (Self-Improvement)
January 1, 2023 – January 1, 2024
Public model hubs have rapidly expanded Generative AI—here’s a selection of what I’ve explored so far: Gaussian-Head Avatars - To enhance user experience with AI-Agents, researched on real-time gaussian head generation accompanied with audio-to-expression. - Created 3D avatars (gaussian-heads) from single image. - Animated avatars from audio-to-expression model outputs. - Animated avatars with expressions extracted from reference videos. - Created singing/talking avatar videos as birthday present. Video Generation - Experimented with quantized video generator models. - Improved generation performance with better image seeding and prompt-engineering. - Created birthday-present videos. Image Generation - Utilized diffusers to create (logo-like) images for personal web portfolio. - Applied text-to-image, image-to-image and inpainting approaches to create and conceptually refine images. - Utilized CLIP/BLIP image interrogation to refine text prompts for improved image generation. - Built a gradio-based image editor to create image masks. - Developed image processing tools for background removal and custom cropping (e.g. rectangle/circle crops with transparent backgrounds etc.) Text Generation - Applied multi-modal (image/video-to-text) approaches for text recognition and visual question answering for key information extraction. - Experimented with conversational LLM models. - Utilized LLMs for summarization by following several approaches (i.e. map-reduce, refiner). - Worked on RAG improvements (for a smart assistant) via document structure understanding (e.g. section extraction, table extraction etc.) Voice Generation - Experimented with text-to-speech (TTS) models as an initial research for a future personal assistant. - Applied voice conversion via pre-embedded public speaker models. - Utilized Hubert embedding to create custom speaker embeddings. - Utilized emotion vectors for better generation control.
Self-Coach (Self-Improvement)
January 1, 2023 – January 1, 2023
Self-Coach is a mobile app to keep track of habits and goals for continuos self-development. The technical goal is to experience DDD and TDD for mobile development with Dart, Flutter and Bloc. Essential know-how: - Advanced testing (TDD, dependency inversion, async/sync tests, mocking, static/dynamic fixtures, integration tests) - Separated domain, infrastructure, application and presentation logic. - Enabled adapters/plug-ins for repositories and ucoqs through interfaces and dependency inversion. - Determined and developed value-objects, entities and aggregations along with validators. - Advanced state management via Bloc (also compared with pure providers). - Ensured state persistence (e.g. for settings etc.) through hydrated blocs. - Managable and composable themes through re-usable theme element plugins (i.e. compositions of colors, icons, typography, shapes) - Ensured robust transaction management through CQRS and failure-handling mechanisms. - Developed on-device backend including complex period-based analytics (requiring SQL recursive queries etc.).
Spiderman (Self-Improvement)
January 1, 2023 – January 1, 2023
Designed and developed a web scraper manager to automate custom scrapper generation, along with customizable context generators (e.g. agents, proxies, devices etc.), bot-detector by-passing etc. Built discoverable registries for composable parts (i.e. contexts, devices, proxies, spiders) through Meta-Classes and AST-parsing. Ensured safe parsing, serialization and processing through Rust-based Pydantic. Advanced development quality with test-driven development (TDD), static-checkers (mypy) and typing. Added automated by-passing features for infinite scroll pages with pop-ups and more-buttons.
Pre-eclampsia Diagnosis (Self-Improvement)
January 1, 2022 – January 1, 2022
Conducted systematic experiments and evaluation to diagnose pre-eclampsia. Applied Explainable AI to aid decision-making process and to ensure transparency on model predictions.
Re-energization Time Prediction
January 1, 2022 – January 1, 2022
When a grid fault occurs, it is crucial to estimate the required time to repair. Customer satisfaction relies on predicting re-energization time. Leadership & Management: Stakeholder management, project management Data Engineering: Designed & developed data pipelines for ingestion and processing. Data Science: Conducted statistical analysis over existing data. Developed classical ML solutions to predict the re-energization time and improved prediction quality. Cross-checked the results with field expectations per damage-cause. Yielded important insights for operations teams for process enhancement.
Smart-Inspect (IN-LEAD 2022)
January 1, 2022 – January 1, 2023
Sabancı Corporation executes IN-LEAD competition among its companies to encourage innovation and coherence. Each company forms a team around a specific problem and the team presents the outcome of their solution to C-level executives of Sabancı Corporation. For solar fields and grid inspection, designed an advanced process enabling enhanced inventory coverage with higher accuracy with less labor effort/time along with reduced risk of work & safety. Award: IN-LEAD 2022 1st place Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, roadmapping, project management Data Engineering: Designed & developed image pipelines for a variety of data sources (e.g. NAS, REST Services) ML Engineering: Developed thermal image analysis solution for solar panel inspection through a variety of methods (e.g. CNNs, classical ML, statistics). Developed object detection based inventory fault detection mechanism for electricty grid operations. Led domain-experts for benchmark dataset construction.
DevOps (Self-Improvement)
January 1, 2021 – January 1, 2023
Service Mesh and Orchestration - Configured Kubernetes (minikube for PoC) and Istio. - Configured a private (docker) registry on the mesh, as a sample use-case. - Configured and tested traffic routing and security features provided by Istio. Containerized Development - Containerized development is crucial for environment management, reproducibility, and deployment. - Configured docker engine and integrated to VSCode for remote container development (both for - Linux & WSL2). Enabled secure remote machine access through SSH and key-based authentication. - Configured hardware access (i.e. GPU etc.), volume management and resource management (e.g. max-allowed memory etc.) through docker-compose. - Supported hot-reload and multi-service deployment.
Apollon (IN-LEAD 2021)
January 1, 2021 – January 1, 2022
Sabancı Corporation executes IN-LEAD competition among its companies to encourage innovation and coherence. Each company forms a team around a specific problem and the team presents the outcome of their solution to C-level executives of Sabancı Corporation. Apollon forsees electricity faults and prioritizes maintenance visits to heal effectively so that electricity outages are prevented, maintenance operations are optimized and enhanced use of investment budget is ensured. Award: IN-LEAD 2021 2nd place Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, roadmapping, project management Data Engineering: Designed & developed data pipelines for a variety of data sources (i.e. SAS, spreadsheets). ML Engineering: Developed predictive maintenance solution for the electricity grid. Proved the effectiveness of the solution both on existing data and on field tests.
Enhanced Missing Data Prediction
January 1, 2021 – January 1, 2022
In an electricity distribution grid, maintaining critical inventory (e.g. transformers etc.) data is crucial for continuous electricity supply and for optimizing a variety of processes (e.g. maintenance, repairment, investment etc.). Due to several reasons (e.g. hand-over areas, outsource quality etc.), some useful properties of inventories may be missing or misleading. Conducted an exhaustive domain analysis with domain-experts and stakeholders. Determined informative inventory properties for predicting required properties. Conducted automated experiments over a large number of inventory types (and their required properties). Auto-reported results along with Explainable AI analysis and confidence-scoring for enhanced decision-making. Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, roadmapping, project management Data Engineering: Designed & develop automated data ingestion and processing pipelines. Constructed a domain specific language (DSL) to define task blueprints and to convert them to respective ingestors and processors (framework abstraction, plug-in based components, auto-sql generation) ML Engineering: Developed domain-specific plug-ins on in-house built ML-Automation solution. Conducted a large number of automatized domain analysis, experiments and reporting. Ensured reliability through confidence-scoring and explainable AI.
Customizable Transformers (Self-Improvement)
January 1, 2021 – January 1, 2021
Paper to code development of Attention is All You Need paper. This project aim to comprehend transformer architecture and to develop customizable software engineered modules; along with presenting detailed guides for transformers and Keras. Model training and decoding can be defined over a DSL; and therefore execution is automated. Abstracts training basics (checkpointing, artifact management etc.) from custom trainers. Such abstraction (for training basics, generators, preprocessors, decoding, callbacks etc.) allows definite interfaces which eases automatization. Custom layers are re-usable and clear. Keeps definite interfaces for generators feeding the models. Users may provide their custom generators to integrate any data source, by inheriting these interfaces. Thus, it separates the ingestion logic from ML/DL architectures/models. Many interesting/hard (Keras) problems are solved: Parameter tying, tying embedders and projector altogether, handling save/load case, parametrized call usage to allow a layer to behave conditionally, domain-agnostic training with inner-outer generators, model and callback serialization, multi-head attention on same tensor, positional encoding - interleaving trick, padding/additional masking when computing attention, custom loss with custom padding mask, subword embeddings, beam search decoding.
Damage-Payment Analysis
January 1, 2021 – January 1, 2021
Energy distribution company should evaluate customer applications and pay the customer if properties of customers got damaged due to electricity distribution faults. ML-supported decision-making system reduces the time-required to process the applications and ensures response time stays in regulation defined boundaries. Leadership & Management: Stakeholder management, project management Data Engineering: Designed & developed data pipelines for ingestion and processing. Data Science: Conducted domain analysis and determined informative features with domain experts. Developed classical ML models for the classification task. Applied appropriate evaluation strategies for scarce data and supported with Explainable AI.
ML-Automation Framework
January 1, 2020 – January 1, 2021
ML-Automation is key for robustness, manageability, re-producability, extendibility, adaptability and time-to-market. However, utilizing commercial products brings many constraints along with high license costs. Instead, applying ML/software engineering rather than scripting leads to ML-Automation which can execute by a simple domain specific language (DSL). Leadership & Management: Roadmapping, project management Data Engineering: Designed & developed automated data ingestion and processing pipelines. Constructed a domain specific language (DSL) to define task blueprints and to convert them to respective ingestors and processors (framework abstraction, plug-in based components) ML Engineering: Designed & developed ML-Automation solution which can execute domain analysis, experimentation and reporting by converting a simple blueprint (defined via DSL) into ML code.
Customs Enforcement (IPA)
January 1, 2020 – January 1, 2020
IPA is financial and technical assistance funding for the candidate and potential candidate countries to help them attain European Union standards. Customs Enforcements Project aims to strengthen the customs surveillance and control function of the Ministry of Commerce throughout the Turkish Customs Territory. Led the project on behalf of Havelsan. Contributed to scoping of the project. Led development of solutions on customs surveillance. Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, partner management, roadmapping, project management Data Engineering: Developed map matching solutions (mapping GPS to roads) and post processing pipelines. ML Engineering: Led development of spatio-temporal embedding solutions both on land and sea.
Customer Segmentation
January 1, 2020 – January 1, 2020
Cogeneration systems produce electricity and heat where they are consumed, minimizing transmission and distribution losses. For cogeneration market, it is crucial to determine the market segments to increase the change of deals, customer satisfaction and operations efficiency. Leadership & Management: Stakeholder management, project management Data Engineering: Designed & developed data pipelines for ingestion and processing. Data Science: Conducted domain analysis and determined informative features with domain experts. To segment data, developed clustering models appropriate for mixed data types (both categorical and numeric features).
Categorical & Time-series Embedders (Self-Improvement)
January 1, 2020 – January 1, 2020
Designed and developed transformer based categorical embedder and time-series embedder. Abstracts shared training and embedding tasks so that developers can focus only on the model architectures. Keeps definite interfaces for generators feeding the models. Users may provide their custom generators to integrate any data source, by inheriting these interfaces. Thus, it separates the ingestion logic from ML/DL architectures/models. Supports 2 main approaches: Seq2seq learning: - Re-usable and generic encoder-decoders allowing choice of recurrent cells, their parameters; the number of layers in encoder and decoder etc. VAE learning: - Enables both continuous and discrete/concerete latent embeddings. - Enables users to define encoder and decoder over a config file for high flexiblity. Provides discriminative wrappers for both approaches. Provides basic processors and layers in addition to the ones in Keras.
IoT Automation
January 1, 2019 – January 1, 2020
As an extension to digital twin project (ITEA3), built IoT pipeline automation module. Led the team to create a PoC IoT kit and find partners to field test the IoT Kit & pipeline automation. Automated sensor add/removal, sensor software updates, stream creation/close and stream analytics. Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, partner management, roadmapping, project management Data Engineering: Automated streaming pipelines from sensor to visualization (ESP32, Raspberry Pi, Mosquitto, Kafka, Spark, InfluxDB, pipe-managing services, Grafana).
Digital Twin (ITEA3 - DayTiMe)
January 1, 2019 – January 1, 2020
DayTiMe is an ITEA3 project with 4 countries and 18 companies. The goal is to create digital representations for physical devices for product enhancements and predictive maintenance. Assigned to lead and develop the project on behalf of Havelsan. Negotiated the project scope and budget with the partners to increase the value and to minimize redundant effort. Ensured effective execution through re-usable modules and automation, under scarce resources. Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, partner management, roadmapping, project management Data Engineering: Developed pipelines on top of in-house built pipeline automation solution (ARAKAT). ML Engineering: Led development of predictive maintenance solutions for use-cases (i.e. medical cabinets) provided by the partners.
ARAKAT - Auto Pipeline Generator
January 1, 2018 – January 1, 2019
Transformed an unattended R&D project into an effective re-usable solution, under scarce resources (both in time and personnel). By promoting the MVP, received international & in-company funds; along with completing the corresponding R&D project with success. Initial version of the solution (ARAKAT) is published as open source which was one of the first attempts in a Turkish defense company. Presented ARAKAT as the core developer at International Confrence on Big Data Deep Learning and Fighting Cyber Terrorism (IBIGDELFT). Leadership & Management: Team-building, leading by example, coaching & mentorship, stakeholder management, roadmapping, project management Data Engineering: Designed domain-specific language (DSL) for auto pipeline generator. Developed DSL interpreter to convert pipline definition blueprints into Spark and Airflow codes, both for batch and stream analytics. Moreover, enabled composition of ML pipelines and SQL-like queries. Ensured high composability, extendibility and customizability of the solution.
R&D Scouts
January 1, 2017 – January 1, 2020
Apart from our main R&D projects, we also designed fast-paced R&D projects for interns and candidate engineers to expand team know-how on a wider range of areas. This enabled the team to be prepared for new opportunities (i.e. immediate demo requests) and for new product features. Building a team from scratch and leading 3-4 weeks projects to completion advanced leadership skills of team members. Hence, in crisis momoments, we were flexible to build any available sub-teams; with minimal alignment costs and advanced pressure management. Such projects also led to an enhanced recruitment process as they created the environment to evaluate the alignment with the team values, in addition to providing the required pre-training. Coaching & mentoring the new generation was quite a fulfilling community service for our team. Essentials to pass to the next generation: team-values, value-orientedness, solution-orientedness, communication, authenticity, resiliency... Graphitage & The Librarian: Knowledge graph to manage and search on highlights of read documents (e.g. for an academic paper: contributions, datasets, libraries, highlights, abstract, keywords, references etc.). Extended with document similarity scoring, topic modeling and search. Pose Similarity Estimation: Pose estimation and similarity scoring on images and videos. Scene Understanding: Image annotation tool (RESTful Web Services, Swagger, Spring Boot, React) Object localization/recognition based scene description and search (Deep Learning) Speech Understanding: Speech collection/annotation tool (RESTful Web Services, Swagger, Spring Boot, React) Language, gender, age, attitude detection over speech (Deep Learning) Virtual Reality: WebVR for dashboards of business intelligence tool (A-frame, Gear VR, Leap Motiton) Resource Monitoring: A tool for monitoring database query performance (Logstash, Elasticsearch, Kibana)
Intelligence Solution (V2)
January 1, 2017 – January 1, 2019
An intelligence solution which started as a project, but turned into a product which targets huge volume and variety of data, in intelligence domain. Extended capabilities of the intelligence product with computer vision, NLP and knowledge graph based analytics. Leadership & Management: Team-building, leading by example, team-branding, customer management Data Engineering: Designed and developed knowledge graph (on ElasticSearch) to enable text-based search, time-range search and location search. Integrated NER-based indexing with search query generation. Developed a vector search solution for image similiarity search and clustering. Designed & developed automated query generator for the knowledge graph. Distributed ML models with Spark. Mapped NER-locations to coordinates via street map service ML Engineering: Developed deep learning based Named Entity Recognition (NER) solution. Led web scraping and annotation of NER dataset for custom entities. Led OCR development for PDF documents. Developed face detection, recognition and embedding solutions. Developed image similarity search and clustering solutions. Software Engineering: Developed services for graph query generator and ML/DL models. Developed UI screens for query generation.
Intelligence Solution (V1)
January 1, 2016 – January 1, 2018
An intelligence solution which started as a project, but turned into a product which targets huge volume and variety of data, in intelligence domain. Equally contributed to data architecture research, design and development in a great team of engineers. Leadership & Management: Team-building, leading by example, team-branding, customer management Data Engineering: Designed domain specific language (DSL) for enabling the customers to define batch and stream analytics in a simple language. Researched on graph computing frameworks and databases (i.e. TinkerPop, Gremlin, Neo4j, Titan etc.). along with Hadoop Ecosystem (i.e. HBase, Hive, Phoenix). Developed parts of the DSL interpreter to transform user query into pipeline code. Software Engineering: Developed a re-usable UI for graph query composer. Extended cytoscape with high-level features.
Active Learning on Big Data (Self-Improvement)
January 1, 2015 – January 1, 2015
Designed and developed a feasible approach to apply machine learning algorithms with polynomial-time complexity to Big Data (with Hadoop, Spark), via active learning.
Music Critic (Self-Improvement)
January 1, 2014 – January 1, 2015
An application for a musician friend. Lets musicians to manage his/her critics and ratings on artists, albums and songs. It allows user-defined analysis on the data via a user-friendly interface.
Optimization via Graph Cuts (Self-Improvement)
January 1, 2014 – January 1, 2014
Implemented Push-Relabel algorithm and its variants with heuristics. Utilized graph cuts in optimization algorithms (e.g. α-expansion, α-β swap) for energy minimization. Built applications for: - Foreground Extraction - Image Restoration - Stereo Imagery
Object Detection (Self-Improvement)
January 1, 2014 – January 1, 2014
Applied Adaboost training to generate a set of cascaded classifiers for face detection (Viola-Jones Detection). Utilized GPU in order to accelerate the training module.
Conditional Random Fields (CRFs) (Self-Improvement)
January 1, 2014 – January 1, 2015
Implemented Conditional Random Fields (CRFs) and its variants (Skip-chain CRFs, Dynamic CRFs). Developed both exact belief propagation and piece-wise training for CRFs. Applied CRFs to: - Noun-phrase Chunking - Information Extraction from e-mails
Haptic-Collaboration Case Study
January 1, 2014 – January 1, 2014
Applied Active Scene Learning to another time-series data, Haptic Collaboration, to learn haptic interaction patterns in dyadic joint object manipulation. Extended the proposed feature extraction method (presented in the literature), to make it more generic and safer.
Active Scene Learning
January 1, 2012 – January 1, 2013
Proposed an active learning strategy for scene segmentation and demonstrated its effectiveness by an extensive empirical evaluation. Introduced a new active learning framework for applications requiring explicit segmentation and applied it to sketched scene segmentation. Publications: Yanık, E., & Sezgin, T. M. (2019). Active Scene Learning. ArXiv:1903.02832 [Cs] Yanık, E., & Sezgin, T. M. (2022). Active Sketch Scene Learning. Available at SSRN 4084576.
Active Learning for Sketch Recognition
January 1, 2011 – January 1, 2012
Active Learning (AL) is a machine learning strategy that aims to reduce the labeling effort by selecting the most informative samples from a pool of unlabeled data. Despite its theoretical appeal, recent empirical results show that active learning does not always yield the expected benefits in practical real world problems. The question is what are the factors effecting AL performance and how. Conducted an extensive empirical analysis to discover the factors and how they effect AL performance. A set of carefully designed experiments and a battery of accompanying statistical tests resulted in a roadmap for robust application of active learning. Publications: Yanık, E., & Sezgin, T. M. (2015). Active learning for sketch recognition. Computers & Graphics, 52, 93-105. Yanık, E., & Sezgin, T. M. Active learning for sketch recognition. Expressive 2016 [Previously published paper track].
Prostate Cancer Diagnosis
January 1, 2011 – January 1, 2011
Conducted a set of experiments to compare the performance of classical ML algorithms on prostate cancer diagnosis.
Selected School Projects
January 1, 2010 – January 1, 2011
Creativity, communication, perseverance, resilience and engineering skills sprout from challenging projects. Observing the pathways from past to now sharpens our coaching & mentroship skills as well as . Wizard of Hunger An application to determine the best meal ordering choice for a given period of time in order to maximize the pleasure by utilizing previous ordering data. (Dynamic programming and HMMs) Electrical Circuit Recognition An application to segment and recognize electrical circuit elements from a sketched electrical circuit. Animation Builder An application to create sketch-based animations with a user-friendly interface. It allows object-based design such that hierarchical objects and behaviors can be formed. It supports animating backgrounds, 3D perception, altering visibility and speech commands. (OpenGL) Extended the application with Microsoft SpeechAPI to enable animation following user-told story. Face Animation Builder An application which captures facial movements of user by a camera, and maps these movements to a sketched face to obtain an animating face. Salsa Guy Animation An application to create a solid human body from geometric objects. Genrated animations with the body such as salsa dancing, ball-playing, walking etc. (OpenGL)
Cultural Fit Analysis
The candidate's extensive personal projects demonstrate a strong drive for continuous learning and self-improvement, which aligns well with a culture of innovation. The diverse range of projects, from DevOps to advanced AI/ML applications, indicates a broad intellectual curiosity. However, the target role is 'Data Analyst', while the candidate's experience leans heavily towards Data Science and ML Engineering. While these skills are valuable, the direct alignment with core data analyst responsibilities (e.g., reporting, dashboarding, specific business intelligence tools) is less explicit, which might require some adaptation.
Soft Skills & Operational Fit
The candidate's project descriptions highlight strong leadership, team-building, coaching, mentorship, stakeholder management, and project management skills. These indicate a high operational fit for senior roles requiring collaboration and strategic execution. The focus on self-improvement projects also suggests a proactive and continuous learning mindset.