Resume

Resume

Emirhan YILDIRIM

Machine Learning EngineerM.Sc. StudentComputer Vision · LLM/RAG Systems · MLOps

Summary

Machine Learning Engineer and M.Sc. student specializing in Computer Vision, LLM/RAG systems, and MLOps. Built a medical blood-cell classifier with inference-time domain adaptation (preprint, live demo), a RAG-grounded project assistant deployed publicly, and a streaming ML pipeline on 10 GB of sensor data at 81 windows/sec. Covers the full stack: training, containerized API (Docker + FastAPI), CI/CD (GitHub Actions). Open to Machine Learning, Computer Vision and AI Engineer roles, remotely or hybrid / on-site near Sakarya. Available part-time or as an intern alongside my M.Sc. studies.

Core Competencies

  • Machine Learning & Computer Vision: PyTorch, TensorFlow, Keras, Scikit-Learn, OpenCV, LightGBM; CNNs, domain adaptation, OOD generalization, inference-time adaptation, Grad-CAM explainability
  • LLM Systems & RAG: ChromaDB, Sentence Transformers, RAG pipeline design, hallucination suppression, Groq API, LLM tool integration
  • MLOps & Production Deployment: Docker (multi-stage builds), GitHub Actions CI/CD, FastAPI, Flask, Hugging Face Spaces, ADWIN drift detection
  • Backend & Data Engineering: Python, SQL, FastAPI, Flask, Dask, Pandas, NumPy, REST APIs
  • Systems & Reliability: Fault-tolerant pipeline design, multi-threaded architectures, real-time telemetry, rate limiting, local test harness development

Experience

Software Engineer Intern | Esbi Bilişim ve Telekomünikasyon

Jul 2026 - Aug 2026 | Sakarya, Türkiye

  • Developed an enterprise CRM web application using ASP.NET Core MVC and Entity Framework Core with Code First migrations (source).
  • Implemented role-based authentication and authorization (RBAC) using ASP.NET Core Identity for secure access control.
  • Built customer and inventory management modules with full CRUD operations, multi-field search, and server-side filtering.
  • Designed a responsive UI with Bootstrap 5 and Razor Views; applied N-Tier architecture with a clean service layer abstraction.

Software Engineer & Founder | zidorun

Jun 2025 - Present | Sakarya, Türkiye

  • Founded and engineered backend infrastructure for a physical collectibles business, handling ~1,000 transactions at a 95%+ success rate with real-time telemetry and error monitoring.
  • Built Listing Pilot: a modular C2C marketplace automation engine with JSON-profiled UI selectors that isolate platform changes.
  • Engineered spatial element-matching algorithms (pairing titles to prices via pixel coordinate proximity) and dynamic overlap detection to prevent obscured taps.
  • Implemented fault-tolerant execution with dynamic retry handling, multi-ID element fallbacks, and adaptive swipe logic for unstable UI states; cut daily manual operations by 90%+ with Telegram Bot alerts.

Software Engineer Intern | Platech Coating

Jun 2025 - Jul 2025 | Sakarya, Türkiye

  • Architected Portal Cleaner Ultimate: a modular desktop RPA suite (Python, Selenium, Tkinter) automating complex ERP workflows.
  • Engineered a multi-threaded GUI that stays responsive during heavy Selenium scraping and XML/XLSX/TXT feed parsing.
  • Built a custom local mock ERP environment (test harness) simulating dynamic UI states and server exceptions for offline end-to-end testing.
  • Impact: Reduced daily processing cycle times from 4-6 hours to under 30 minutes (90%+ reduction); the system remained in active production use.

Research & Projects

Computer Vision & Machine Learning Research Project

Aug 2026 - Present | Sakarya University

  • Ongoing supervised research, continued as my M.Sc. thesis from Sep 2026; developing real-time video processing methods with a focus on low latency and temporal consistency (details and results withheld at my supervisor’s request until publication).
  • Designing the quantitative evaluation: perceptual image-quality metrics, statistical significance testing, and effect sizes.
  • Studying how video enhancement and processing methods tuned on one data domain generalize to others.

Computer Vision for Blood Cell Microscopy

Sep 2026 - Present | Sakarya University

  • Supervised research project building an image processing toolkit for stained blood smear microscopy, under my advisor’s supervision.
  • Engineered the pipeline from raw microscope images to quantitative cell measurements: physics-based colour and stain handling, cross-slide stain normalization, nucleus and cell segmentation, and texture and shape feature extraction.
  • Validated segmentation and quantitative extraction against expert clinical annotations using Python, PyTorch, OpenCV, and Scikit-Learn.

RAG Project Assistant: Hybrid-Source Q&A System

Jun 2026 | GitHub

  • Engineered code extraction with Python’s ast module instead of regex, giving accurate docstrings and function signatures without false positives on multi-line strings.
  • Calibrated a similarity-threshold gate (ChromaDB, L2 < 1.40) after verifying that standard embeddings under-scored table-heavy markdown; it suppresses hallucination on out-of-scope queries.
  • Resolved cross-project context leakage (shared “API” terminology causing retrieval contamination) with keyword-based project detection and metadata filtering, falling back to full-corpus retrieval.
  • Deployed as a FastAPI service on Hugging Face Spaces (Docker) with Groq Llama 3.3 70B inference, IP-based rate limiting (slowapi), and a public JS chat widget.

Scalable Kinematic Action Recognition: Industry 5.0

Apr 2026 - May 2026 | Sakarya University | GitHub

  • Built out-of-core ingestion with Dask and engineered 528 sliding-window kinematic features from 10 GB of motion-capture sensor data (132 channels, 15 action classes).
  • Unsupervised baseline (MiniBatchKMeans + PCA + Hungarian matching) reached 0.14 Macro F1 on the 15 classes; a supervised RandomForest on the binary subset reached 0.9995 Macro F1, using a systematic every-5th-window split so 50%-overlapping windows cannot leak between train and test.
  • Designed real-time streaming with Adaptive Random Forest + ADWIN: 81 windows/sec, 59 MB peak RAM, and drift from a simulated label-flip fault detected within 58 windows (~290 ms).
  • Kaggle variant: kinematic feature engineering (elbow joint angles, bilateral distances, rotation norms) + LightGBM/RandomForest soft-voting ensemble, 0.94169 private leaderboard accuracy.

WBC Analyzer: AI-Powered Pathology Assistant

Jan 2026 - Apr 2026 | Sakarya University | GitHub | Live demo

  • Achieved 98.53% in-distribution accuracy on the Raabin-WBC dataset with a custom DenseNet121, WBCAttentionBlock, and MedSwish activation.
  • Raised OOD accuracy from 56.96% to 89.05% (+32.09 pp) on unseen hardware via a 5-step Medical Enhanced Filter (MEF) preprocessing pipeline and inference-time domain adaptation (Reinhard normalization, TTA, binary routing; the OOD set holds two classes), with no retraining.
  • Handled severe class imbalance with a custom WBC Focal Loss combined with Label Smoothing.
  • Containerized with a multi-stage Docker build; GitHub Actions CI/CD (pytest + image build on push); Flask REST API at 200-350 ms inference with a /health observability endpoint.
  • Implemented a multimodal LLM observer agent (GPT-4o & Gemini 2.5 Flash) that interprets Grad-CAM heatmaps to flag shortcut learning, decoupled from the core inference pipeline.
  • Published preprint on ResearchGate: doi.org/10.13140/RG.2.2.34201.79208

Popcorn Wagon: Hybrid Movie Recommender System

Jan 2025 - May 2025 | GitHub

  • Built a hybrid recommendation engine combining content-based filtering (TMDB API) and collaborative filtering (MovieLens + SVD + Annoy).
  • Implemented millisecond-level similarity search with Spotify’s Annoy and scalable ETL pipelines using Pandas and Dask.
  • Developed a full-stack Flask application with user authentication, personalized movie lists, tagging, and SQLAlchemy-based session management.

Education

M.Sc., Information Systems Engineering
Sakarya University | Sep 2026 - Feb 2028
Focus and thesis research: Medical AI, Computer Vision, Machine Learning

B.Sc., Information Systems Engineering
Sakarya University | Sep 2021 - Sep 2026

Languages

  • Turkish: Native
  • English: Professional working proficiency (YOKDIL: 91.25/100, Aug 2026)

Certifications

  • Machine Learning Specialization - DeepLearning.AI / Stanford University, Coursera (Aug 2026)
  • CS50x: Introduction to Computer Science - Harvard University (May 2025)
  • Odoo Student Diploma - Odoo S.A. (Dec 2025)
  • Courses completed (DeepLearning.AI): Unsupervised Learning, Recommenders, Reinforcement Learning (Aug 2026); Advanced Learning Algorithms (Jul 2026); Supervised Machine Learning: Regression and Classification (Jun 2026)

Publications

  • Yildirim, E. (2026). “Achieving Robust Out-of-Distribution Generalization in Peripheral Blood Smears via Custom Attention Mechanisms, Medical Enhanced Filtering, and Inference-Time Domain Adaptation.” Preprint, ResearchGate. DOI

References

Available upon request.