Open to opportunities

Tuna Kömürcü

@tunakmrc

Message

AI engineer at Mikrogrup who developed MomTest AI and improved GİB CAPTCHA recognition accuracy from 41% to 93%.

Turkey

Message

At Mikrogrup, I developed MomTest AI end-to-end, building an agent loop with Function Calling and Structured Outputs, plus guardrails against leading questions in automated customer-discovery interviews. I also owned GİB CAPTCHA OCR model training and fine-tuning, raising recognition accuracy from 41% to 93%.

At Başarı Mobile, I designed and built a database-agnostic Text2SQL Agent, including a Qdrant-based hybrid retrieval system for dynamic schema selection. It reached about 95% accuracy on simple queries and 85% on medium-complexity queries.

At Havelsan Technology Radar, I designed and trained an Edge-Attentive Residual CNN to denoise raw radar imagery. My projects also include Voxel, a real-time AI voice-agent pitch coach, and an eight-stage multi-agent system for legal and public tender analysis.

Experience

Work history, roles, and key accomplishments

AI Engineering Intern / Developer

Mikrogrup

May 2026 - Aug 2026 (3 months)

Developed MomTest AI end-to-end, an agent loop using Function Calling and Structured Outputs with guardrails to prevent leading-question bias. Integrated Markdown-based rule files, deployed on AWS, and built test suites; also fine-tuned GİB CAPTCHA OCR, raising accuracy from 41% to 93%.

Function Calling Structured Outputs AWS EC2 S3 IAM Cloudwatch Unit Testing Integration Testing Model Fine Tuning OCR TROCR

AI Engineering Intern / Developer

Başarı Mobile

Sep 2025 - Jan 2026 (4 months)

Designed and built a database-agnostic Text2SQL Agent end-to-end, including architecture and prompt engineering. Built a Hybrid Retrieval (Dense + BM25) architecture with Qdrant for dynamic schema retrieval, reaching ~95% accuracy on simple and ~85% on medium-complexity queries; load-tested with Locust and monitored latency via Prometheus and Grafana.

Text2SQL Prompt Engineering Hybrid Retrieval Dense Retrieval BM25 Qdrant HNSW Locust Prometheus Grafana Response Time Monitoring

AI Engineering Intern

Havelsan Technology Radar

Sep 2024 - Jan 2025 (4 months)

Designed and trained an Edge-Attentive Residual CNN from scratch, using attention mechanisms to denoise raw radar imagery. Implemented hybrid activation and loss functions, applied ensemble learning, and managed the end-to-end training pipeline.

CNN Attention Mechanisms Image Denoising Ensemble Learning Machine Learning Pipelines Activation Functions Loss Functions

Education

Degrees, certifications, and relevant coursework

TOBB University of Economics and Technology

Bachelor of Science, Artificial Intelligence Engineering

2021 - 2026

Activities and societies: Graduation Project: Stack Underflow - AI Legal & Tender Analyst; developed an 8-stage multi-agent LLM system with RAG pipelines for public tender analysis and antitrust legal workflows.

Pursuing a Bachelor of Science in Artificial Intelligence Engineering with a focus on Machine Learning, Deep Learning, NLP, and Software Architecture.