Project Showcase and Technical Summaries

This page is organized for quick screening with a consistent format:

• Duration
• Tech
• Impact
• Links

Project Cards

Computer Vision & Machine Learning Research Project

• Duration: Aug 2026 - Present
• Tech: Computer Vision, Video Processing, Digital Image Processing, Applied Machine Learning
• Impact:
  • Supervised research at Sakarya University, continued as my M.Sc. thesis from Sep 2026.
  • Real-time video processing methods with a focus on low latency and temporal consistency.
  • Quantitative evaluation design: perceptual image-quality metrics, statistical testing, and effect sizes.
  • Details and results are withheld at my supervisor’s request until publication.

Computer Vision for Blood Cell Microscopy

• Duration: Sep 2026 - Present
• Tech: Python, PyTorch, Scikit-Learn, OpenCV
• Impact:
  • Image processing toolkit for stained blood smear microscopy, from raw microscope images to quantitative cell measurements.
  • Physics-based colour and stain handling, cross-slide stain normalization, nucleus and cell segmentation, texture and shape features.
  • Validated against expert annotations, under my advisor’s supervision.

ASP.NET Core MVC CRM Application

• Duration: Jul 2026 - Aug 2026 (Esbi Bilişim ve Telekomünikasyon)
• Tech: ASP.NET Core MVC, Entity Framework Core (Code First), ASP.NET Core Identity, Bootstrap 5, Razor Views, SQL, C#
• Impact:
  • Role-based authentication and authorization with ASP.NET Core Identity.
  • Customer and inventory management modules with full CRUD, search, and filtering.
  • N-Tier architecture with a service layer abstraction and a responsive Bootstrap 5 UI.
• Links:
  • GitHub

WBC Analyzer: AI-Powered Pathology Assistant

• Duration: Jan 2026 - Apr 2026
• Tech: DenseNet121, WBCAttention, MedSwish, MEF Preprocessing, Flask REST API, Grad-CAM, GPT-4o, Gemini 2.5 Flash
• Impact:
  • Achieved 98.53% in-distribution accuracy and 89.05% OOD accuracy on unseen hardware data (+32.09 pp over unadapted baseline, retraining-free; the OOD set holds two classes, so binary routing applies).
  • Designed 5-step Medical Enhanced Filter (MEF) pipeline for cross-device staining and exposure normalization.
  • Integrated autonomous multi-modal LLM agent (GPT-4o / Gemini fallback) for Grad-CAM-based shortcut detection.
  • Published academic preprint (DOI: 10.13140/RG.2.2.34201.79208).
• Links:
  • Case Study
  • GitHub
  • Live Demo

Kinematic Action Recognition: Industry 5.0 Pipeline

• Duration: Apr 2026 - May 2026
• Tech: Python, Dask, Scikit-Learn, River (ARF + ADWIN), Parquet, LightGBM, RandomForest
• Impact:
  • Processed 10 GB proprietary motion-capture CSV into Parquet (3× compression) via out-of-core Dask pipeline.
  • Extracted 97,612 sliding-window feature vectors (528 features) from 9.7M sensor rows.
  • RandomForest binary classifier: Macro F1 ≈ 0.9995 on the binary subset, 1.3s train, 8.7 MB RAM.
  • Adaptive Random Forest streaming: 81 windows/sec, drift detected in ~58 windows (~290 ms) via ADWIN.
  • Kaggle variant with a LightGBM + RandomForest soft-voting ensemble reached 0.94169 private leaderboard accuracy.
• Links:
  • Case Study
  • GitHub

RAG Project Assistant: Hybrid-Source Q&A System

• Duration: Jun 2026
• Tech: sentence-transformers, ChromaDB, Groq Llama 3.3 70B, FastAPI, Docker, slowapi
• Impact:
  • Built a Retrieval-Augmented Generation system answering questions about five personal ML/automation projects, combining AST-extracted code structure with hand-curated documentation in a hybrid knowledge base.
  • Engineered a similarity-threshold gate (ChromaDB, L2 distance < 1.40, empirically calibrated) paired with a strict grounding prompt requiring explicit source attribution to suppress hallucination on out-of-scope queries.
  • Deployed as a FastAPI service on Hugging Face Spaces (Docker), migrating inference from local Ollama to Groq’s Llama 3.3 70B; added IP-based rate limiting (slowapi) and a public JS chat widget embedded on portfolio site.
• Links:
  • GitHub

Popcorn Wagon: Hybrid Movie Recommender System

• Duration: Jan 2025 - May 2025
• Tech: TruncatedSVD, TMDB API, MovieLens, Pandas, Dask, Spotify Annoy
• Impact:
  • Built a hybrid recommendation strategy combining collaborative and content-based methods.
  • Processed large datasets with a scalable ETL pipeline.
  • Reduced recommendation latency to millisecond level using approximate nearest neighbors.
• Links:
  • GitHub

CV Download

• PDF CV: Please contact me at emirhan0yildirim@gmail.com for the latest PDF version.

Professional Experience Highlights

Software Engineer Intern - Esbi Bilişim ve Telekomünikasyon (Jul 2026 - Aug 2026)

• Developed a CRM web application using ASP.NET Core MVC and Entity Framework Core with Code First migrations.
• Implemented role-based authentication and authorization using ASP.NET Core Identity.
• 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 service layer abstraction.

Software Engineer & Founder - zidorun (Jun 2025 - Present)

• 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, config-driven C2C marketplace automation engine with JSON-profiled UI selectors that isolate platform changes.
• Implemented fault-tolerant execution with dynamic retry handling and Telegram Bot alerts, cutting daily manual operations by 90%+.

Software Engineer Intern - Platech Coating (Jun 2025 - Jul 2025)

• Architected Portal Cleaner Ultimate: a modular desktop RPA suite (Python, Selenium, Tkinter) automating complex ERP workflows.
• Designed a multi-threaded system architecture with fault-tolerant retry loops and logging, ensuring stable data extraction and parsing of XML, XLSX, and TXT feeds.
• Built a local mock ERP server environment (test harness) to simulate server-side network dropouts and exception states for offline validation.
• Impact: Reduced daily data processing cycle times by 90% (saving 4–6 hours of manual entry down to under 30 minutes); system remained in active production use post-internship.

Detailed case-study posts with architecture diagrams and API snippets will be published next.