E
TREN
Open to AI & Backend Opportunities

AI & Backend Developer

I bring AIto life throughsecure backendarchitectures.

I bring AI and computer vision models to life through secure, scalable backend architectures.

I am a Computer Engineering graduate from Gazi University. I focus on connecting AI models and data infrastructure with working, end-to-end backend systems. Beyond writing code, I care about understanding the real problem, planning the architecture, and validating what I build with technical metrics.

GitHubLinkedInMailContact

Computer Vision

Backend Systems

LLM Products

Search & Data

Scroll

01

About

Who am I?

During my undergraduate studies and internships, I focused on integrating artificial intelligence, computer vision, and data systems with backend applications. As a part-time developer at the Gazi Institute of Informatics, I worked on face recognition and exam scheduling projects. During my internships at TRT, Ramsa Defence, and Neokod R&D, I contributed to OCR, LLM feasibility, and large-scale data indexing work. My goal is to take theoretical knowledge beyond demo level and turn it into working products that cover database design, API security, and model serving. My volunteer work with the Gazi AI Community has also strengthened my appreciation for team coordination and a collaborative engineering culture.

02

Expertise

AI & Computer Vision

01

I build real-time detection, recognition, tracking, and decision-support systems with YOLO, OpenCV, PyTorch, and tracking algorithms.

YOLOOpenCVPyTorchObject Tracking

Backend Systems

02

I design readable, maintainable, and extensible API, data-model, service-boundary, and integration layers.

FastAPIPostgreSQLMongoDBREST

LLMs & Product Development

03

I turn local and cloud LLM workflows into secure, resilient, and observable product architectures.

OllamaLlamaRAGAgents

Data & Search Infrastructure

04

I design systems that deliver fast, meaningful, and traceable results through Elasticsearch, data cleaning, and analytics pipelines.

ElasticsearchPandasKibanaAnalytics

03

Projects

01KariyerKoçAn AI-powered career coaching platform that combines CV analysis, ATS simulation, interview practice, job search, and cover-letter generation.View detailsClose details
LLM Product2026io.net Hackathon 2025 — 1st Place
My role

Built the full-stack product, FastAPI service architecture, LLM integration, CV/ATS/interview modules, and Docker-based runtime.

Architecture

React + TypeScript frontend -> FastAPI REST API -> PostgreSQL -> io Intelligence Llama 3.3 70B, Groq Whisper, Edge-TTS, and JobSpy integrations.

Problem

Job seekers had to manage CV quality, role fit, interview preparation, and application materials across disconnected tools.

Solution

Modularized LLM analysis services behind a FastAPI backend and unified CV upload, analysis, interview, job-search, and cover-letter flows in one React experience.

Outcome

Won first place at the io.net Hackathon 2025.
Developed 57 endpoints across authentication, CV, analysis, interview, voice, chatbot, job-search, and cover-letter modules.
Configured the PostgreSQL, FastAPI, and React stack to start with one Docker Compose command.
FastAPIReactTypeScriptPostgreSQLDockerLlama 3.3 70BGroq WhisperEdge-TTSJobSpy
GitHub

01 / 03

Project media

02AI WorkbenchA local AI workbench for macOS that unifies chat, source-grounded batch generation, model and RAM management, and controlled coding workflows.View detailsClose details
Local AI Platform2026
My role

Designed the product scope and security boundaries; built the Tauri/Rust backend, React UI, SQLite job engine, Ollama integration, source retrieval pipeline, and Aider workflow.

Architecture

React/TypeScript UI -> typed Tauri command boundary -> Rust services -> SQLite WAL + FTS5 -> Ollama; a single-AI resource scheduler coordinates chat, batch, and Aider roles based on model and RAM state.

Problem

Processing thousands of educational records with local LLMs required reliable recovery, restricted file and shell access, and safe scheduling of chat, batch, and coding models within 16 GB of memory.

Solution

Built a resilient per-record transactional batch engine. The pipeline losslessly chunks course notes, retrieves relevant local sources, and supports pilot approval, retry/resume, human review, and traceable source IDs.

Outcome

Unified streaming local chat, source-grounded batch generation, and controlled Aider sessions in one desktop experience.
Completed per-record transactions, retry, pause/resume, crash recovery, pilot approval, and JSONL/CSV/Markdown export flows.
Implemented memory protection that unloads verified models under pressure and resumes batch work once the system returns to a 30% safety margin.
RustTauri 2ReactTypeScriptSQLiteFTS5OllamaQwenAiderxterm.js
03DroDetA real-time detection system that finds and tracks small drone targets from camera or video input and stabilizes threat decisions over time.View detailsClose details
Computer Vision2026
My role

Worked on YOLO training and fine-tuning, TrackBoost decision stabilization, the PyQt interface, diagnostic reporting, and MQTT alert flow.

Architecture

Video/Camera -> YOLOv26s-P2 -> Temperature Scaling -> Duplicate Merge -> BIoUTracker -> TrackBoost -> Threat Evaluator -> PyQt UI + MQTT alerts.

Problem

Distant drones could shrink to 10–20 pixels. The previous YOLOv8s model and frame-level decision logic could switch the same target between drone and bird, creating unstable operator alerts.

Solution

After an initial YOLOv8s dataset of roughly 30K images, moved the pipeline to YOLOv26s-P2 and applied two fine-tuning stages. Added Temperature Scaling, Duplicate Merge, and TrackBoost so product decisions are made at track level rather than from raw single-frame output.

Outcome

Measured 0.928 precision, 0.887 recall, 0.914 mAP50, and 0.608 mAP50-95 on the YOLOv26s-P2 FT2 validation set.
Reported 0.976 drone AP and 0.918 all-class mAP@0.5 in BoxPR evaluation.
PythonYOLOv26s-P2PyTorchOpenCVBIoUTrackerTrackBoostPyQtMQTTDocker
GitHubDemo

01 / 01

Project media

04YDS KelimeA Flutter learning app for YDS/YÖKDİL exams that combines 6,100+ vocabulary cards, thousands of exercises, and personalized review scheduling.View detailsClose details
Mobile EdTech2026
My role

Built the entire product independently, from concept and data model to content validation, Flutter UI, learning algorithm, Supabase infrastructure, and Android release preparation.

Architecture

Flutter/Riverpod UI -> Drift/SQLite local database and offline use -> planner + FSRS learning engine -> Supabase Auth, PostgreSQL, RLS, and cross-device synchronization.

Problem

Learners often study large vocabulary lists without context and cannot track what to review, daily goals, and genuine progress in one place.

Solution

Combined definitions, examples, translations, synonyms/antonyms, word families, and usage patterns with multiple question types. A goal-based planner controls new-word load while FSRS schedules each review from individual performance.

Outcome

Made more than 6,100 searchable vocabulary cards and thousands of exercises available in one app.
Translated active recall, spaced repetition, and progressive difficulty research into a goal-based planner and FSRS product flow.
Completed Google sign-in, account deletion, RLS isolation, local-first use, cross-device sync, and Android release signing.
FlutterDartRiverpodDriftSQLiteFSRSSupabasePostgreSQLGoogle OAuth
05DDSA full-stack prototype that classifies uploaded images as authentic or deepfake.View detailsClose details
Deepfake Detection2025
My role

Built the MobileNetV2 model integration, backend inference service, React interface, and Docker Compose runtime.

Architecture

React frontend -> Python API -> PyTorch MobileNetV2 inference -> Docker Compose.

Problem

Because image manipulation is difficult to identify manually, uploaded images needed a fast and understandable classification flow.

Solution

Connected the PyTorch model to an API and displayed upload state, classification, confidence, and processing time in the frontend.

Outcome

Reported approximately 90% validation accuracy in the README.
Kept model weights, inference service, training pipeline, and Docker Compose configuration in the repository.
PythonPyTorchMobileNetV2OpenCVScikit-learnFlaskFastAPIReactDocker
06EXAMBEAn exam management system that solves course, student, session, and room constraints with CP-SAT for university scheduling workflows.View detailsClose details
Optimization2026
My role

Contributed across the entire co-developed project: product direction, system architecture, FastAPI backend, CP-SAT engine, React/TypeScript UI, and document generation.

Architecture

React/Vite frontend -> FastAPI backend -> SQLModel/SQLite -> CP-SAT scheduling engine -> document generator.

Problem

Student conflicts, session load, and room capacity had to be managed simultaneously during exam scheduling.

Solution

Unified Excel import, data management, CP-SAT scheduling, and Word/Excel document generation in one system while leading product direction and architecture.

Outcome

The system is actively used by the Gazi Institute of Informatics.
Built scheduling, data import, and document-generation modules across roughly 45 files and 5,800 backend/frontend lines.
Implemented the CP-SAT constraint model for student load, session balance, and room capacity.
FastAPIReactTypeScriptGoogle OR-ToolsCP-SATSQLModelSQLitePandaspython-docx
07NeoNurseA search infrastructure for equivalent-drug discovery and drug–drug interaction queries over Turkish medication data.View detailsClose details
Search & Data2026
My role

Built the Excel/CSV ingestion, MongoDB data model, Elasticsearch search pipeline, and system architecture documentation.

Architecture

Excel/CSV import -> MongoDB collections -> Elasticsearch index -> equivalent and interaction query flow.

Problem

Equivalent-drug and interaction data was fragmented across sources and needed a fast, queryable structure.

Solution

Structured TİTCK medication data, ATC/active-ingredient fields, and DrugBank DDI data for a MongoDB/Elasticsearch pipeline.

Outcome

Cataloged 23,045 medication records, 114 DrugBank interaction types, and thousands of drug pairs.
Based equivalent-drug matching on ATC code and active-ingredient alignment.
PythonPandasMongoDBElasticsearchKibanaDockerDrugBank DDIATC
08Video2TextA GPU-accelerated OCR prototype that reads text from video frames and exports it with timestamps.View detailsClose details
OCR Pipeline2026
My role

Built frame extraction, the PaddleOCR wrapper, JSON/TXT output, and logging layers.

Architecture

Video input -> FrameExtractor -> PaddleOCRWrapper -> JsonWriter + SimpleRowWriter -> grouped TXT output.

Problem

Manual extraction of text from video was time-consuming; frame-level OCR was needed with timestamped output.

Solution

Extracted frames at configured intervals with OpenCV, applied Turkish OCR with PaddleOCR 3.1.0, and wrote results to JSON and TXT.

Outcome

Delivered an end-to-end pipeline for frame extraction, OCR inference, JSON logging, and grouped TXT export.
Managed Turkish language support, angle classification, and PP-OCRv5 model configuration inside the PaddleOCR wrapper.
PythonOpenCVPaddleOCRPaddlePaddle GPUNumPyJSON
09Face Recognition SystemA desktop security system that records personnel entry and exit events across two cameras using face recognition and tracking.View detailsClose details
Computer Vision2025
My role

Contributed throughout direction, integration, and implementation; worked on face detection, recognition, person tracking, the desktop UI, and event logging.

Architecture

Entry/Exit cameras -> SCRFD -> ArcFace R100 -> BYTETracker -> PyQt UI -> SQLite + Excel reporting.

Problem

Personnel had to be recognized in real time across two camera streams, tracked across frames, and logged consistently for entry and exit.

Solution

Combined SCRFD face detection, ArcFace R100 identity matching, and BYTETracker in one PyQt application, logging events to SQLite and exporting Excel reports.

Outcome

Unified personnel enrollment, live entry/exit monitoring, event logs, daily backups, and Excel reporting in one desktop application.
PythonOpenCVPyTorchSCRFDArcFaceBYTETrackerPyQt5SQLitePandas
10Employee Attrition PredictionA data-mining study that predicts employee attrition risk on the IBM HR dataset while accounting for class imbalance.View detailsClose details
Data Mining2026
My role

Contributed throughout the co-developed project, focusing on feature engineering, leakage-safe evaluation, sampler/model comparison, and decision-threshold optimization.

Architecture

IBM HR data -> feature engineering -> preprocessing + sampler pipeline -> model comparison -> OOF threshold selection -> test and robustness analysis.

Problem

Employees who left represented only 16.12% of the dataset, so high-accuracy models could still miss most at-risk employees.

Solution

Applied sampling only within training folds and compared model/SMOTE-family combinations through out-of-fold predictions and recall/F1 tradeoffs.

Outcome

The final XGB + SMOTEENN recall model achieved 80.27% accuracy, 42.86% precision, 70.21% recall, and 53.23% F1.
PythonPandasScikit-learnImbalanced-learnXGBoostSMOTEENNOptunaMatplotlib
11B2B Process AutomationA desktop prototype that structures public company data and unifies B2B research with user-controlled outreach preparation.View detailsClose details
Workflow Automation2025
My role

Built the Electron UI, browser automation, Excel data pipeline, company email-pattern generation, and delivery-service integration.

Architecture

Electron UI -> Node.js/Puppeteer research flow -> Excel/JSON pipeline -> Python format generation -> Brevo API.

Problem

Company research, contact-list preparation, and corporate email-pattern generation required repetitive manual work across separate tools.

Solution

Processed user-provided company lists through browser automation, structured public company information in Excel, and presented possible corporate email formats for user review.

Outcome

Combined company lists, employee data, email-pattern preparation, and outreach setup in a repeatable desktop workflow.
ElectronNode.jsPuppeteerPythonPandasExcelJSBrevo API
12TraSimA Java simulation that runs the Nagel–Schreckenberg traffic model sequentially and in parallel, producing performance and traffic-behavior analysis.View detailsClose details
Parallel Computing2026
My role

In the co-developed project, my teammate built the web UI while I designed the model and algorithms, sequential and parallel simulators, validation tests, benchmarks, congestion analysis, and reports.

Architecture

NaSch model core -> sequential/ExecutorService simulators -> validation and benchmarks -> CSV/Markdown reports + web dashboard.

Problem

Parallelization had to preserve vehicle count, collision-free movement, and deterministic equality with sequential results.

Solution

Partitioned the circular road into cells, preserved synchronous steps with double buffering, and verified sequential/parallel equivalence through automated tests.

Outcome

Reported up to 5.214× speedup with eight threads in one-million-cell benchmark scenarios.
Java 17Java ThreadsExecutorServiceWeb WorkersLeafletChart.jsOpenStreetMap
13BLE AlarmAn IoT prototype that sends an ESP32 alarm input to an Android app over BLE and delivers a local notification.View detailsClose details
IoT Prototype2025
My role

Built ESP32 BLE communication, Android client connectivity, local notifications, and Telegram alert flow.

Architecture

Alarm sensor -> ESP32 BLE server -> Android BluetoothGatt client -> local notification + Telegram Bot API.

Problem

A physical alarm trigger needed to reach a nearby phone through a low-cost wireless connection.

Solution

Published alarm-pin changes as BLE characteristic notifications on the ESP32, then listened on Android and generated local and Telegram alerts.

Outcome

Built an end-to-end prototype that carries the alarm signal from the ESP32 to the Android client and notifies the user.
ESP32Arduino C++Bluetooth Low EnergyAndroidJavaTelegram Bot API

04

Experience / Education

September 2021 - June 2026

B.Sc. in Computer Engineering

Gazi University

Focused on artificial intelligence, computer vision, and backend architectures, translating academic foundations into practical projects.

November 2024 - June 2026

Part-time Software Developer

Gazi Institute of Informatics

Led development work and team coordination for a face-recognition entry/exit control system. Also contributed actively to the software and coordination of EXAMBE, which optimizes exam scheduling workflows.

January 2026 - February 2026

Software Development Intern

Neokod R&D Software

Migrated large-scale catalog data to MongoDB and automated Python-based data transfer and indexing for Elasticsearch. Conducted feasibility studies for LLM integrations.

June 2025 - August 2025

Software Development Intern

TRT & Ramsa Defence

At TRT, developed a Python/OCR video transcription application. At Ramsa Defence, worked on software modules and API designs for defense-industry projects using Java and Python.

05

Technology

AI / Vision

PythonPyTorchYOLOOpenCVSCRFDArcFaceByteTrack

Backend

FastAPIRustREST APIPostgreSQLMongoDBSQLiteDocker

Data / Search

PandasElasticsearchKibanaScikit-learnXGBoost

Frontend / Product

ReactTypeScriptTauriAstroTailwindElectron

Contact

Let's talk about an AI, backend, or product-focused role.

This portfolio will continue to grow with technical evidence and detailed project case studies. For now, the fastest way to reach me is through one of my active profiles.

GitHubLinkedInMail